2004-11-22 00:37:56 +00:00
|
|
|
/*
|
2005-11-05 19:57:48 +00:00
|
|
|
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
|
|
|
|
|
* University Research and Technology
|
|
|
|
|
* Corporation. All rights reserved.
|
|
|
|
|
* Copyright (c) 2004-2005 The University of Tennessee and The University
|
|
|
|
|
* of Tennessee Research Foundation. All rights
|
|
|
|
|
* reserved.
|
2006-02-07 03:32:36 +00:00
|
|
|
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
|
2004-11-28 20:09:25 +00:00
|
|
|
* University of Stuttgart. All rights reserved.
|
2005-03-24 12:43:37 +00:00
|
|
|
* Copyright (c) 2004-2005 The Regents of the University of California.
|
|
|
|
|
* All rights reserved.
|
2007-05-31 18:27:54 +00:00
|
|
|
* Copyright (c) 2007 Sun Microsystems, Inc. All rights reserved.
|
2004-11-22 01:38:40 +00:00
|
|
|
* $COPYRIGHT$
|
2006-02-07 03:32:36 +00:00
|
|
|
*
|
2004-11-22 01:38:40 +00:00
|
|
|
* Additional copyrights may follow
|
2006-02-07 03:32:36 +00:00
|
|
|
*
|
2004-11-22 00:37:56 +00:00
|
|
|
* $HEADER$
|
|
|
|
|
*/
|
|
|
|
|
|
2006-02-12 01:33:29 +00:00
|
|
|
#include "orte_config.h"
|
2008-02-28 01:57:57 +00:00
|
|
|
#include "orte/constants.h"
|
|
|
|
|
#include "orte/types.h"
|
2006-11-03 18:55:05 +00:00
|
|
|
|
2004-11-20 19:12:43 +00:00
|
|
|
#include <string.h>
|
2006-11-03 18:55:05 +00:00
|
|
|
#ifdef HAVE_SYS_TIME_H
|
|
|
|
|
#include <sys/time.h>
|
|
|
|
|
#endif /* HAVE_SYS_TIME_H */
|
2004-11-20 19:12:43 +00:00
|
|
|
|
2005-10-31 16:21:51 +00:00
|
|
|
#include "opal/threads/condition.h"
|
2005-07-03 23:31:27 +00:00
|
|
|
#include "opal/util/output.h"
|
2006-11-28 00:06:25 +00:00
|
|
|
#include "opal/util/bit_ops.h"
|
2006-11-03 18:55:05 +00:00
|
|
|
|
2008-03-05 22:44:35 +00:00
|
|
|
#include "opal/class/opal_hash_table.h"
|
2006-02-12 01:33:29 +00:00
|
|
|
#include "orte/util/proc_info.h"
|
2008-02-28 01:57:57 +00:00
|
|
|
#include "opal/dss/dss.h"
|
2006-02-12 01:33:29 +00:00
|
|
|
#include "orte/mca/errmgr/errmgr.h"
|
Bring in the code for routing xcast stage gate messages via the local orteds. This code is inactive unless you specifically request it via an mca param oob_xcast_mode (can be set to "linear" or "direct"). Direct mode is the old standard method where we send messages directly to each MPI process. Linear mode sends the xcast message via the orteds, with the HNP sending the message to each orted directly.
There is a binomial algorithm in the code (i.e., the HNP would send to a subset of the orteds, which then relay it on according to the typical log-2 algo), but that has a bug in it so the code won't let you select it even if you tried (and the mca param doesn't show, so you'd *really* have to try).
This also involved a slight change to the oob.xcast API, so propagated that as required.
Note: this has *only* been tested on rsh, SLURM, and Bproc environments (now that it has been transferred to the OMPI trunk, I'll need to re-test it [only done rsh so far]). It should work fine on any environment that uses the ORTE daemons - anywhere else, you are on your own... :-)
Also, correct a mistake where the orte_debug_flag was declared an int, but the mca param was set as a bool. Move the storage for that flag to the orte/runtime/params.c and orte/runtime/params.h files appropriately.
This commit was SVN r14475.
2007-04-23 18:41:04 +00:00
|
|
|
#include "orte/mca/odls/odls_types.h"
|
|
|
|
|
#include "orte/mca/rml/rml.h"
|
2008-02-28 01:57:57 +00:00
|
|
|
#include "orte/runtime/orte_globals.h"
|
|
|
|
|
#include "orte/util/name_fns.h"
|
|
|
|
|
#include "orte/orted/orted.h"
|
2008-02-28 19:58:32 +00:00
|
|
|
#include "orte/runtime/orte_wait.h"
|
2004-11-20 19:12:43 +00:00
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
#include "orte/mca/grpcomm/base/base.h"
|
2007-07-20 01:34:02 +00:00
|
|
|
#include "grpcomm_basic.h"
|
2004-11-20 19:12:43 +00:00
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
/* Local functions */
|
2007-07-20 01:34:02 +00:00
|
|
|
static int xcast_binomial_tree(orte_jobid_t job,
|
2008-02-28 01:57:57 +00:00
|
|
|
opal_buffer_t *buffer,
|
2007-07-20 01:34:02 +00:00
|
|
|
orte_rml_tag_t tag);
|
2006-12-01 22:30:39 +00:00
|
|
|
|
2007-07-20 01:34:02 +00:00
|
|
|
static int xcast_linear(orte_jobid_t job,
|
2008-02-28 01:57:57 +00:00
|
|
|
opal_buffer_t *buffer,
|
2007-07-20 01:34:02 +00:00
|
|
|
orte_rml_tag_t tag);
|
Bring in the code for routing xcast stage gate messages via the local orteds. This code is inactive unless you specifically request it via an mca param oob_xcast_mode (can be set to "linear" or "direct"). Direct mode is the old standard method where we send messages directly to each MPI process. Linear mode sends the xcast message via the orteds, with the HNP sending the message to each orted directly.
There is a binomial algorithm in the code (i.e., the HNP would send to a subset of the orteds, which then relay it on according to the typical log-2 algo), but that has a bug in it so the code won't let you select it even if you tried (and the mca param doesn't show, so you'd *really* have to try).
This also involved a slight change to the oob.xcast API, so propagated that as required.
Note: this has *only* been tested on rsh, SLURM, and Bproc environments (now that it has been transferred to the OMPI trunk, I'll need to re-test it [only done rsh so far]). It should work fine on any environment that uses the ORTE daemons - anywhere else, you are on your own... :-)
Also, correct a mistake where the orte_debug_flag was declared an int, but the mca param was set as a bool. Move the storage for that flag to the orte/runtime/params.c and orte/runtime/params.h files appropriately.
This commit was SVN r14475.
2007-04-23 18:41:04 +00:00
|
|
|
|
2007-07-20 01:34:02 +00:00
|
|
|
static int xcast_direct(orte_jobid_t job,
|
2008-02-28 01:57:57 +00:00
|
|
|
opal_buffer_t *buffer,
|
2007-07-20 01:34:02 +00:00
|
|
|
orte_rml_tag_t tag);
|
2008-04-09 22:10:53 +00:00
|
|
|
static int find_parent(int rank, int parent, int me, int num_procs,
|
|
|
|
|
int *num_children, opal_list_t *children);
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/* Local global variables */
|
|
|
|
|
static orte_process_name_t my_parent;
|
|
|
|
|
static opal_list_t *my_children;
|
|
|
|
|
static int my_num_children;
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
|
2008-03-24 20:50:31 +00:00
|
|
|
/* Static API's */
|
|
|
|
|
static int init(void);
|
|
|
|
|
static void finalize(void);
|
|
|
|
|
static int xcast(orte_jobid_t job,
|
|
|
|
|
opal_buffer_t *buffer,
|
|
|
|
|
orte_rml_tag_t tag);
|
2008-04-09 22:10:53 +00:00
|
|
|
static int allgather(opal_buffer_t *sbuf, opal_buffer_t *rbuf);
|
|
|
|
|
static int barrier(void);
|
|
|
|
|
static opal_list_t* next_recips(orte_grpcomm_mode_t mode);
|
|
|
|
|
static int daemon_collective(orte_jobid_t jobid,
|
|
|
|
|
orte_std_cntr_t num_local_contributors,
|
|
|
|
|
orte_grpcomm_coll_t type,
|
|
|
|
|
opal_buffer_t *data,
|
2008-04-30 19:49:53 +00:00
|
|
|
bool hnp_has_local_procs);
|
2008-04-09 22:10:53 +00:00
|
|
|
static int update_trees(void);
|
2008-03-24 20:50:31 +00:00
|
|
|
|
|
|
|
|
/* Module def */
|
|
|
|
|
orte_grpcomm_base_module_t orte_grpcomm_basic_module = {
|
|
|
|
|
init,
|
|
|
|
|
finalize,
|
|
|
|
|
xcast,
|
2008-04-09 22:10:53 +00:00
|
|
|
allgather,
|
2008-03-24 20:50:31 +00:00
|
|
|
orte_grpcomm_base_allgather_list,
|
2008-04-09 22:10:53 +00:00
|
|
|
barrier,
|
|
|
|
|
daemon_collective,
|
|
|
|
|
update_trees,
|
2008-03-24 20:50:31 +00:00
|
|
|
next_recips,
|
|
|
|
|
orte_grpcomm_base_set_proc_attr,
|
|
|
|
|
orte_grpcomm_base_get_proc_attr,
|
|
|
|
|
orte_grpcomm_base_modex,
|
|
|
|
|
orte_grpcomm_base_purge_proc_attrs
|
|
|
|
|
};
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Initialize the module
|
|
|
|
|
*/
|
|
|
|
|
static int init(void)
|
|
|
|
|
{
|
|
|
|
|
int rc;
|
|
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
/* setup the local global variables */
|
|
|
|
|
if (orte_process_info.hnp || orte_process_info.daemon) {
|
|
|
|
|
update_trees();
|
|
|
|
|
}
|
|
|
|
|
|
2008-03-24 20:50:31 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_grpcomm_base_modex_init())) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Finalize the module
|
|
|
|
|
*/
|
|
|
|
|
static void finalize(void)
|
|
|
|
|
{
|
2008-04-09 22:10:53 +00:00
|
|
|
opal_list_item_t *item;
|
|
|
|
|
|
|
|
|
|
if (orte_process_info.hnp || orte_process_info.daemon) {
|
|
|
|
|
/* deconstruct the child list */
|
|
|
|
|
while (NULL != (item = opal_list_remove_first(my_children))) {
|
|
|
|
|
OBJ_RELEASE(item);
|
|
|
|
|
}
|
|
|
|
|
OBJ_RELEASE(my_children);
|
|
|
|
|
my_num_children = 0;
|
|
|
|
|
}
|
|
|
|
|
|
2008-03-24 20:50:31 +00:00
|
|
|
orte_grpcomm_base_modex_finalize();
|
|
|
|
|
}
|
|
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
static int update_trees(void)
|
|
|
|
|
{
|
|
|
|
|
opal_list_item_t *item;
|
|
|
|
|
|
|
|
|
|
if (NULL != my_children) {
|
|
|
|
|
while (NULL != (item = opal_list_remove_first(my_children))) {
|
|
|
|
|
OBJ_RELEASE(item);
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
my_children = OBJ_NEW(opal_list_t);
|
|
|
|
|
}
|
|
|
|
|
my_num_children = 0;
|
|
|
|
|
my_parent.jobid = ORTE_PROC_MY_NAME->jobid;
|
|
|
|
|
my_parent.vpid = find_parent(0, 0, ORTE_PROC_MY_NAME->vpid,
|
|
|
|
|
orte_process_info.num_procs,
|
|
|
|
|
&my_num_children, my_children);
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic update trees found %d children num_procs %d",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
my_num_children, orte_process_info.num_procs));
|
|
|
|
|
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
2008-03-24 20:50:31 +00:00
|
|
|
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
/**
|
|
|
|
|
* A "broadcast-like" function to a job's processes.
|
|
|
|
|
* @param jobid The job whose processes are to receive the message
|
|
|
|
|
* @param buffer The data to broadcast
|
|
|
|
|
*/
|
2006-12-01 22:30:39 +00:00
|
|
|
|
2007-07-20 01:34:02 +00:00
|
|
|
static int xcast(orte_jobid_t job,
|
2008-02-28 01:57:57 +00:00
|
|
|
opal_buffer_t *buffer,
|
2007-07-20 01:34:02 +00:00
|
|
|
orte_rml_tag_t tag)
|
2006-12-01 22:30:39 +00:00
|
|
|
{
|
2006-12-03 00:19:11 +00:00
|
|
|
int rc = ORTE_SUCCESS;
|
2006-12-01 22:30:39 +00:00
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
2008-03-17 19:34:36 +00:00
|
|
|
"%s grpcomm:xcast sent to job %s tag %ld",
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
2008-02-28 01:57:57 +00:00
|
|
|
ORTE_JOBID_PRINT(job), (long)tag));
|
2007-07-20 01:34:02 +00:00
|
|
|
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
/* if there is no message to send, then just return ok */
|
|
|
|
|
if (NULL == buffer) {
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
2008-03-17 19:34:36 +00:00
|
|
|
"%s grpcomm:xcast: num_procs %ld linear xover: %ld binomial xover: %ld",
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
2008-03-03 20:07:02 +00:00
|
|
|
(long)orte_process_info.num_procs,
|
|
|
|
|
(long)orte_grpcomm_basic.xcast_linear_xover,
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
(long)orte_grpcomm_basic.xcast_binomial_xover));
|
2008-04-09 22:10:53 +00:00
|
|
|
|
|
|
|
|
if (orte_abnormal_term_ordered) {
|
|
|
|
|
/* We insist that the direct xcast mode be used when
|
When we can detect that a daemon has failed, then we would like to terminate the system without having it lock up. The "hang" is currently caused by the system attempting to send messages to the daemons (specifically, ordering them to kill their local procs and then terminate). Unfortunately, without some idea of which daemon has died, the system hangs while attempting to send a message to someone who is no longer alive.
This commit introduces the necessary logic to avoid that conflict. If a PLS component can identify that a daemon has failed, then we will set a flag indicating that fact. The xcast system will subsequently check that flag and, if it is set, will send all messages direct to the recipient. In the case of "kill local procs" and "terminate", the messages will go directly to each orted, thus bypassing any orted that has failed.
In addition, the xcast system will -not- wait for the messages to complete, but will return immediately (i.e., operate in non-blocking mode). Orterun will wait (via an event timer) for a period of time based on the number of daemons in the system to allow the messages to attempt to be delivered - at the end of that time, orterun will simply exit, alerting the user to the problem and -strongly- recommending they run orte-clean.
I could only test this on slurm for the case where all daemons unexpectedly died - srun apparently only executes its waitpid callback when all launched functions terminate. I have asked that Jeff integrate this capability into the OOB as he is working on it so that we execute it whenever a socket to an orted is unexpectedly closed. Meantime, the functionality will rarely get called, but at least the logic is available for anyone whose environment can support it.
This commit was SVN r16451.
2007-10-15 18:00:30 +00:00
|
|
|
* an orted has failed as we cannot rely on alternative
|
|
|
|
|
* methods to reach all orteds and/or procs
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
*/
|
2007-07-20 01:34:02 +00:00
|
|
|
rc = xcast_direct(job, buffer, tag);
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
goto DONE;
|
2006-12-01 22:30:39 +00:00
|
|
|
}
|
Bring in the code for routing xcast stage gate messages via the local orteds. This code is inactive unless you specifically request it via an mca param oob_xcast_mode (can be set to "linear" or "direct"). Direct mode is the old standard method where we send messages directly to each MPI process. Linear mode sends the xcast message via the orteds, with the HNP sending the message to each orted directly.
There is a binomial algorithm in the code (i.e., the HNP would send to a subset of the orteds, which then relay it on according to the typical log-2 algo), but that has a bug in it so the code won't let you select it even if you tried (and the mca param doesn't show, so you'd *really* have to try).
This also involved a slight change to the oob.xcast API, so propagated that as required.
Note: this has *only* been tested on rsh, SLURM, and Bproc environments (now that it has been transferred to the OMPI trunk, I'll need to re-test it [only done rsh so far]). It should work fine on any environment that uses the ORTE daemons - anywhere else, you are on your own... :-)
Also, correct a mistake where the orte_debug_flag was declared an int, but the mca param was set as a bool. Move the storage for that flag to the orte/runtime/params.c and orte/runtime/params.h files appropriately.
This commit was SVN r14475.
2007-04-23 18:41:04 +00:00
|
|
|
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
/* now use the crossover points to select the proper transmission
|
|
|
|
|
* mode. We have built-in default crossover points for this
|
|
|
|
|
* decision tree, but the user is free to alter them as
|
|
|
|
|
* they wish via MCA params
|
|
|
|
|
*/
|
|
|
|
|
|
2008-03-03 20:07:02 +00:00
|
|
|
if (orte_process_info.num_procs < orte_grpcomm_basic.xcast_linear_xover) {
|
2007-07-20 01:34:02 +00:00
|
|
|
rc = xcast_direct(job, buffer, tag);
|
2008-03-03 20:07:02 +00:00
|
|
|
} else if (orte_process_info.num_procs < orte_grpcomm_basic.xcast_binomial_xover) {
|
2007-07-20 01:34:02 +00:00
|
|
|
rc = xcast_linear(job, buffer, tag);
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
} else {
|
2007-07-20 01:34:02 +00:00
|
|
|
rc = xcast_binomial_tree(job, buffer, tag);
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
DONE:
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
|
2006-12-01 22:30:39 +00:00
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
2007-07-20 01:34:02 +00:00
|
|
|
static int xcast_binomial_tree(orte_jobid_t job,
|
2008-02-28 01:57:57 +00:00
|
|
|
opal_buffer_t *buffer,
|
2007-07-20 01:34:02 +00:00
|
|
|
orte_rml_tag_t tag)
|
2004-11-20 19:12:43 +00:00
|
|
|
{
|
2008-02-28 01:57:57 +00:00
|
|
|
orte_daemon_cmd_flag_t command;
|
|
|
|
|
orte_grpcomm_mode_t mode;
|
2007-05-23 14:06:32 +00:00
|
|
|
int rc;
|
2008-02-28 01:57:57 +00:00
|
|
|
opal_buffer_t *buf;
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
2008-03-17 19:34:36 +00:00
|
|
|
"%s grpcomm:entering xcast_binomial",
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
2008-03-03 20:07:02 +00:00
|
|
|
/* binomial xcast can only go through the daemons as app procs are
|
|
|
|
|
* not allowed to relay messages.
|
|
|
|
|
* first, need to pack the msg and be sure to include routing info so it
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
* can properly be sent through the daemons
|
|
|
|
|
*/
|
2008-02-28 01:57:57 +00:00
|
|
|
buf = OBJ_NEW(opal_buffer_t);
|
2007-05-23 14:06:32 +00:00
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
/* tell the daemon to process and relay */
|
|
|
|
|
command = ORTE_DAEMON_PROCESS_AND_RELAY_CMD;
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &command, 1, ORTE_DAEMON_CMD))) {
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
|
|
|
|
|
/* tell the daemon the routing algorithm this xmission is using */
|
|
|
|
|
mode = ORTE_GRPCOMM_BINOMIAL;
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &mode, 1, ORTE_GRPCOMM_MODE))) {
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
/* if this isn't intended for the daemon command tag, then we better
|
|
|
|
|
* tell the daemon to deliver it to the procs, and what job is supposed
|
|
|
|
|
* to get it - this occurs when a caller just wants to send something
|
|
|
|
|
* to all the procs in a job. In that use-case, the caller doesn't know
|
|
|
|
|
* anything about inserting daemon commands or what routing algo might
|
|
|
|
|
* be used, so we have to help them out a little. Functions that are
|
|
|
|
|
* sending commands to the daemons themselves are smart enough to know
|
|
|
|
|
* what they need to do.
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_RML_TAG_DAEMON != tag) {
|
|
|
|
|
command = ORTE_DAEMON_MESSAGE_LOCAL_PROCS;
|
2008-02-28 01:57:57 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &command, 1, ORTE_DAEMON_CMD))) {
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &job, 1, ORTE_JOBID))) {
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &tag, 1, ORTE_RML_TAG))) {
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
/* copy the payload into the new buffer - this is non-destructive, so our
|
|
|
|
|
* caller is still responsible for releasing any memory in the buffer they
|
|
|
|
|
* gave to us
|
|
|
|
|
*/
|
2008-02-28 01:57:57 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.copy_payload(buf, buffer))) {
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
Bring in the code for routing xcast stage gate messages via the local orteds. This code is inactive unless you specifically request it via an mca param oob_xcast_mode (can be set to "linear" or "direct"). Direct mode is the old standard method where we send messages directly to each MPI process. Linear mode sends the xcast message via the orteds, with the HNP sending the message to each orted directly.
There is a binomial algorithm in the code (i.e., the HNP would send to a subset of the orteds, which then relay it on according to the typical log-2 algo), but that has a bug in it so the code won't let you select it even if you tried (and the mca param doesn't show, so you'd *really* have to try).
This also involved a slight change to the oob.xcast API, so propagated that as required.
Note: this has *only* been tested on rsh, SLURM, and Bproc environments (now that it has been transferred to the OMPI trunk, I'll need to re-test it [only done rsh so far]). It should work fine on any environment that uses the ORTE daemons - anywhere else, you are on your own... :-)
Also, correct a mistake where the orte_debug_flag was declared an int, but the mca param was set as a bool. Move the storage for that flag to the orte/runtime/params.c and orte/runtime/params.h files appropriately.
This commit was SVN r14475.
2007-04-23 18:41:04 +00:00
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
|
2008-03-17 19:34:36 +00:00
|
|
|
"%s grpcomm:xcast_binomial: buffer size %ld",
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
(long)buf->bytes_used));
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
/* all we need to do is send this to the HNP - the relay logic
|
2007-07-23 15:00:39 +00:00
|
|
|
* will ensure everyone else gets it!
|
|
|
|
|
*/
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
|
2008-03-17 19:34:36 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:xcast_binomial: sending %s => %s",
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
2008-02-28 01:57:57 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_HNP)));
|
|
|
|
|
|
2008-02-28 19:58:32 +00:00
|
|
|
/* if I am the HNP, just set things up so the cmd processor gets called.
|
|
|
|
|
* We don't want to message ourselves as this can create circular logic
|
|
|
|
|
* in the RML. Instead, this macro will set a zero-time event which will
|
|
|
|
|
* cause the buffer to be processed by the cmd processor - probably will
|
|
|
|
|
* fire right away, but that's okay
|
|
|
|
|
* The macro makes a copy of the buffer, so it's okay to release it here
|
|
|
|
|
*/
|
2008-02-28 01:57:57 +00:00
|
|
|
if (orte_process_info.hnp) {
|
2008-02-28 19:58:32 +00:00
|
|
|
ORTE_MESSAGE_EVENT(ORTE_PROC_MY_NAME, buf, ORTE_RML_TAG_DAEMON, orte_daemon_cmd_processor);
|
2008-02-28 01:57:57 +00:00
|
|
|
} else {
|
2008-04-09 22:10:53 +00:00
|
|
|
/* otherwise, send it to the HNP for relay */
|
2008-02-28 01:57:57 +00:00
|
|
|
if (0 > (rc = orte_rml.send_buffer(ORTE_PROC_MY_HNP, buf, ORTE_RML_TAG_DAEMON, 0))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
rc = ORTE_SUCCESS;
|
|
|
|
|
}
|
2005-10-31 16:21:51 +00:00
|
|
|
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
CLEANUP:
|
2008-02-28 01:57:57 +00:00
|
|
|
OBJ_RELEASE(buf);
|
|
|
|
|
|
2008-03-17 19:34:36 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:xcast_binomial: completed",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
return rc;
|
2004-11-20 19:12:43 +00:00
|
|
|
}
|
|
|
|
|
|
2007-07-20 01:34:02 +00:00
|
|
|
static int xcast_linear(orte_jobid_t job,
|
2008-02-28 01:57:57 +00:00
|
|
|
opal_buffer_t *buffer,
|
2007-07-20 01:34:02 +00:00
|
|
|
orte_rml_tag_t tag)
|
Bring in the code for routing xcast stage gate messages via the local orteds. This code is inactive unless you specifically request it via an mca param oob_xcast_mode (can be set to "linear" or "direct"). Direct mode is the old standard method where we send messages directly to each MPI process. Linear mode sends the xcast message via the orteds, with the HNP sending the message to each orted directly.
There is a binomial algorithm in the code (i.e., the HNP would send to a subset of the orteds, which then relay it on according to the typical log-2 algo), but that has a bug in it so the code won't let you select it even if you tried (and the mca param doesn't show, so you'd *really* have to try).
This also involved a slight change to the oob.xcast API, so propagated that as required.
Note: this has *only* been tested on rsh, SLURM, and Bproc environments (now that it has been transferred to the OMPI trunk, I'll need to re-test it [only done rsh so far]). It should work fine on any environment that uses the ORTE daemons - anywhere else, you are on your own... :-)
Also, correct a mistake where the orte_debug_flag was declared an int, but the mca param was set as a bool. Move the storage for that flag to the orte/runtime/params.c and orte/runtime/params.h files appropriately.
This commit was SVN r14475.
2007-04-23 18:41:04 +00:00
|
|
|
{
|
|
|
|
|
int rc;
|
2008-02-28 01:57:57 +00:00
|
|
|
opal_buffer_t *buf;
|
|
|
|
|
orte_daemon_cmd_flag_t command;
|
2008-03-03 20:07:02 +00:00
|
|
|
orte_grpcomm_mode_t mode;
|
Bring in the code for routing xcast stage gate messages via the local orteds. This code is inactive unless you specifically request it via an mca param oob_xcast_mode (can be set to "linear" or "direct"). Direct mode is the old standard method where we send messages directly to each MPI process. Linear mode sends the xcast message via the orteds, with the HNP sending the message to each orted directly.
There is a binomial algorithm in the code (i.e., the HNP would send to a subset of the orteds, which then relay it on according to the typical log-2 algo), but that has a bug in it so the code won't let you select it even if you tried (and the mca param doesn't show, so you'd *really* have to try).
This also involved a slight change to the oob.xcast API, so propagated that as required.
Note: this has *only* been tested on rsh, SLURM, and Bproc environments (now that it has been transferred to the OMPI trunk, I'll need to re-test it [only done rsh so far]). It should work fine on any environment that uses the ORTE daemons - anywhere else, you are on your own... :-)
Also, correct a mistake where the orte_debug_flag was declared an int, but the mca param was set as a bool. Move the storage for that flag to the orte/runtime/params.c and orte/runtime/params.h files appropriately.
This commit was SVN r14475.
2007-04-23 18:41:04 +00:00
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
2008-03-17 19:34:36 +00:00
|
|
|
"%s grpcomm:entering xcast_linear",
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
/* since we have to pack some additional info into the buffer to be
|
|
|
|
|
* sent to the daemons, we create a new buffer into which we will
|
|
|
|
|
* put the intermediate payload - i.e., the info that goes to the
|
|
|
|
|
* daemon. This buffer will contain all the info needed by the
|
|
|
|
|
* daemon, plus the payload intended for the processes themselves
|
|
|
|
|
*/
|
2008-02-28 01:57:57 +00:00
|
|
|
buf = OBJ_NEW(opal_buffer_t);
|
2008-03-03 20:07:02 +00:00
|
|
|
|
|
|
|
|
/* if we are an application proc, then send this to our HNP so
|
|
|
|
|
* we don't try to talk to every daemon directly ourselves. This
|
|
|
|
|
* is necessary since we don't know how many daemons there are!
|
2008-04-09 22:10:53 +00:00
|
|
|
*
|
|
|
|
|
* Likewise, a daemon who is not the HNP will also let the HNP
|
|
|
|
|
* act as the relay to avoid opening unnecessary connections
|
|
|
|
|
* and rattling messages around the system if daemons are not
|
|
|
|
|
* fully connected
|
2008-03-03 20:07:02 +00:00
|
|
|
*/
|
2008-05-02 14:30:07 +00:00
|
|
|
|
|
|
|
|
/* tell the HNP to relay */
|
|
|
|
|
command = ORTE_DAEMON_PROCESS_AND_RELAY_CMD;
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &command, 1, ORTE_DAEMON_CMD))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
/* tell the HNP the routing algorithm this xmission is using */
|
|
|
|
|
mode = ORTE_GRPCOMM_LINEAR;
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &mode, 1, ORTE_GRPCOMM_MODE))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
2008-03-03 20:07:02 +00:00
|
|
|
}
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
/* if this isn't intended for the daemon command tag, then we better
|
|
|
|
|
* tell the daemon to deliver it to the procs, and what job is supposed
|
|
|
|
|
* to get it - this occurs when a caller just wants to send something
|
|
|
|
|
* to all the procs in a job. In that use-case, the caller doesn't know
|
|
|
|
|
* anything about inserting daemon commands or what routing algo might
|
|
|
|
|
* be used, so we have to help them out a little. Functions that are
|
|
|
|
|
* sending commands to the daemons themselves are smart enough to know
|
|
|
|
|
* what they need to do.
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_RML_TAG_DAEMON != tag) {
|
|
|
|
|
command = ORTE_DAEMON_MESSAGE_LOCAL_PROCS;
|
2008-02-28 01:57:57 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &command, 1, ORTE_DAEMON_CMD))) {
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &job, 1, ORTE_JOBID))) {
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &tag, 1, ORTE_RML_TAG))) {
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* copy the payload into the new buffer - this is non-destructive, so our
|
|
|
|
|
* caller is still responsible for releasing any memory in the buffer they
|
|
|
|
|
* gave to us
|
|
|
|
|
*/
|
2008-02-28 01:57:57 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.copy_payload(buf, buffer))) {
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
|
2008-03-17 19:34:36 +00:00
|
|
|
"%s grpcomm:xcast_linear: buffer size %ld",
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
(long)buf->bytes_used));
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
/* if I am the HNP, just set things up so the cmd processor gets called.
|
|
|
|
|
* We don't want to message ourselves as this can create circular logic
|
|
|
|
|
* in the RML. Instead, this macro will set a zero-time event which will
|
|
|
|
|
* cause the buffer to be processed by the cmd processor - probably will
|
|
|
|
|
* fire right away, but that's okay
|
|
|
|
|
* The macro makes a copy of the buffer, so it's okay to release it here
|
|
|
|
|
*/
|
|
|
|
|
if (orte_process_info.hnp) {
|
|
|
|
|
ORTE_MESSAGE_EVENT(ORTE_PROC_MY_NAME, buf, ORTE_RML_TAG_DAEMON, orte_daemon_cmd_processor);
|
|
|
|
|
} else {
|
|
|
|
|
/* otherwise, send it to the HNP for relay */
|
2008-03-03 20:07:02 +00:00
|
|
|
if (0 > (rc = orte_rml.send_buffer(ORTE_PROC_MY_HNP, buf, ORTE_RML_TAG_DAEMON, 0))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
rc = ORTE_SUCCESS;
|
|
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
|
Bring in the code for routing xcast stage gate messages via the local orteds. This code is inactive unless you specifically request it via an mca param oob_xcast_mode (can be set to "linear" or "direct"). Direct mode is the old standard method where we send messages directly to each MPI process. Linear mode sends the xcast message via the orteds, with the HNP sending the message to each orted directly.
There is a binomial algorithm in the code (i.e., the HNP would send to a subset of the orteds, which then relay it on according to the typical log-2 algo), but that has a bug in it so the code won't let you select it even if you tried (and the mca param doesn't show, so you'd *really* have to try).
This also involved a slight change to the oob.xcast API, so propagated that as required.
Note: this has *only* been tested on rsh, SLURM, and Bproc environments (now that it has been transferred to the OMPI trunk, I'll need to re-test it [only done rsh so far]). It should work fine on any environment that uses the ORTE daemons - anywhere else, you are on your own... :-)
Also, correct a mistake where the orte_debug_flag was declared an int, but the mca param was set as a bool. Move the storage for that flag to the orte/runtime/params.c and orte/runtime/params.h files appropriately.
This commit was SVN r14475.
2007-04-23 18:41:04 +00:00
|
|
|
CLEANUP:
|
2008-03-03 20:07:02 +00:00
|
|
|
/* release the buffer */
|
Bring in the generalized xcast communication system along with the correspondingly revised orted launch. I will send a message out to developers explaining the basic changes. In brief:
1. generalize orte_rml.xcast to become a general broadcast-like messaging system. Messages can now be sent to any tag on the daemons or processes. Note that any message sent via xcast will be delivered to ALL processes in the specified job - you don't get to pick and choose. At a later date, we will introduce an augmented capability that will use the daemons as relays, but will allow you to send to a specified array of process names.
2. extended orte_rml.xcast so it supports more scalable message routing methodologies. At the moment, we support three: (a) direct, which sends the message directly to all recipients; (b) linear, which sends the message to the local daemon on each node, which then relays it to its own local procs; and (b) binomial, which sends the message via a binomial algo across all the daemons, each of which then relays to its own local procs. The crossover points between the algos are adjustable via MCA param, or you can simply demand that a specific algo be used.
3. orteds no longer exhibit two types of behavior: bootproxy or VM. Orteds now always behave like they are part of a virtual machine - they simply launch a job if mpirun tells them to do so. This is another step towards creating an "orteboot" functionality, but also provided a clean system for supporting message relaying.
Note one major impact of this commit: multiple daemons on a node cannot be supported any longer! Only a single daemon/node is now allowed.
This commit is known to break support for the following environments: POE, Xgrid, Xcpu, Windows. It has been tested on rsh, SLURM, and Bproc. Modifications for TM support have been made but could not be verified due to machine problems at LANL. Modifications for SGE have been made but could not be verified. The developers for the non-verified environments will be separately notified along with suggestions on how to fix the problems.
This commit was SVN r15007.
2007-06-12 13:28:54 +00:00
|
|
|
OBJ_RELEASE(buf);
|
2008-03-17 19:34:36 +00:00
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:xcast_linear: completed",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
return rc;
|
Bring in the code for routing xcast stage gate messages via the local orteds. This code is inactive unless you specifically request it via an mca param oob_xcast_mode (can be set to "linear" or "direct"). Direct mode is the old standard method where we send messages directly to each MPI process. Linear mode sends the xcast message via the orteds, with the HNP sending the message to each orted directly.
There is a binomial algorithm in the code (i.e., the HNP would send to a subset of the orteds, which then relay it on according to the typical log-2 algo), but that has a bug in it so the code won't let you select it even if you tried (and the mca param doesn't show, so you'd *really* have to try).
This also involved a slight change to the oob.xcast API, so propagated that as required.
Note: this has *only* been tested on rsh, SLURM, and Bproc environments (now that it has been transferred to the OMPI trunk, I'll need to re-test it [only done rsh so far]). It should work fine on any environment that uses the ORTE daemons - anywhere else, you are on your own... :-)
Also, correct a mistake where the orte_debug_flag was declared an int, but the mca param was set as a bool. Move the storage for that flag to the orte/runtime/params.c and orte/runtime/params.h files appropriately.
This commit was SVN r14475.
2007-04-23 18:41:04 +00:00
|
|
|
}
|
|
|
|
|
|
2008-03-06 18:13:05 +00:00
|
|
|
static int relay_via_hnp(orte_jobid_t job,
|
|
|
|
|
opal_buffer_t *buffer,
|
|
|
|
|
orte_rml_tag_t tag) {
|
|
|
|
|
opal_buffer_t *buf;
|
|
|
|
|
orte_daemon_cmd_flag_t command;
|
|
|
|
|
orte_grpcomm_mode_t mode;
|
|
|
|
|
int rc;
|
|
|
|
|
|
2008-03-17 19:34:36 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm: relaying buffer to HNP",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
2008-03-06 18:13:05 +00:00
|
|
|
/* since we have to pack some additional info into the buffer
|
|
|
|
|
* for this case, we create a new buffer into to contain all the
|
|
|
|
|
* info needed plus the payload
|
|
|
|
|
*/
|
|
|
|
|
buf = OBJ_NEW(opal_buffer_t);
|
|
|
|
|
/* start by telling the HNP to relay */
|
|
|
|
|
command = ORTE_DAEMON_PROCESS_AND_RELAY_CMD;
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &command, 1, ORTE_DAEMON_CMD))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
/* default to the LINEAR mode since this is equivalent to
|
|
|
|
|
* DIRECT for daemons
|
|
|
|
|
*/
|
|
|
|
|
mode = ORTE_GRPCOMM_LINEAR;
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &mode, 1, ORTE_GRPCOMM_MODE))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
/* if the target isn't the daemon tag, then we have to add the proper
|
|
|
|
|
* command so the daemon's know what to do
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_RML_TAG_DAEMON != tag) {
|
|
|
|
|
command = ORTE_DAEMON_MESSAGE_LOCAL_PROCS;
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &command, 1, ORTE_DAEMON_CMD))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &job, 1, ORTE_JOBID))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &tag, 1, ORTE_RML_TAG))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
/* copy the payload into the new buffer - this is non-destructive, so our
|
|
|
|
|
* caller is still responsible for releasing any memory in the buffer they
|
|
|
|
|
* gave to us
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.copy_payload(buf, buffer))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
if (0 > (rc = orte_rml.send_buffer(ORTE_PROC_MY_HNP, buf, ORTE_RML_TAG_DAEMON, 0))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
rc = ORTE_SUCCESS;
|
|
|
|
|
|
|
|
|
|
CLEANUP:
|
|
|
|
|
OBJ_RELEASE(buf);
|
2008-03-17 19:34:36 +00:00
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm: buffer relayed to HNP",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
2008-03-06 18:13:05 +00:00
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
2007-07-20 01:34:02 +00:00
|
|
|
static int xcast_direct(orte_jobid_t job,
|
2008-02-28 01:57:57 +00:00
|
|
|
opal_buffer_t *buffer,
|
2007-07-20 01:34:02 +00:00
|
|
|
orte_rml_tag_t tag)
|
2006-12-01 22:30:39 +00:00
|
|
|
{
|
|
|
|
|
int rc;
|
2008-02-28 01:57:57 +00:00
|
|
|
orte_process_name_t peer;
|
2008-03-24 20:50:31 +00:00
|
|
|
orte_vpid_t i, num_targets=0;
|
2008-03-03 20:07:02 +00:00
|
|
|
opal_buffer_t *buf=NULL, *bfr=buffer;
|
|
|
|
|
orte_daemon_cmd_flag_t command;
|
|
|
|
|
orte_rml_tag_t target=tag;
|
|
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
2008-03-17 19:34:36 +00:00
|
|
|
"%s grpcomm: entering xcast_direct",
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
2008-03-06 18:13:05 +00:00
|
|
|
/* if I am applicaton proc */
|
|
|
|
|
if (!orte_process_info.hnp &&
|
|
|
|
|
!orte_process_info.daemon &&
|
|
|
|
|
!orte_process_info.tool) {
|
|
|
|
|
/* if this is going to some job other
|
|
|
|
|
* than my own, then we have to send it via the HNP as I have
|
|
|
|
|
* no way of knowing how many procs are in the other job.
|
2008-03-03 20:07:02 +00:00
|
|
|
*/
|
2008-03-06 18:13:05 +00:00
|
|
|
if (ORTE_PROC_MY_NAME->jobid != job) {
|
|
|
|
|
if (ORTE_SUCCESS != (rc = relay_via_hnp(job, buffer, tag))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
2008-05-02 14:30:07 +00:00
|
|
|
goto CLEANUP;
|
2008-03-03 20:07:02 +00:00
|
|
|
}
|
2008-03-06 18:13:05 +00:00
|
|
|
/* if it is my jobid, then we can just send this ourselves -
|
|
|
|
|
* set the target tag
|
2008-03-03 20:07:02 +00:00
|
|
|
*/
|
2008-03-06 18:13:05 +00:00
|
|
|
target = tag;
|
|
|
|
|
/* set number of procs to the #procs in our job */
|
|
|
|
|
num_targets = orte_process_info.num_procs;
|
|
|
|
|
/* point to the right buffer */
|
|
|
|
|
bfr = buffer;
|
|
|
|
|
/* go to send it */
|
|
|
|
|
goto SEND;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* if I am a daemon */
|
|
|
|
|
if (orte_process_info.daemon) {
|
|
|
|
|
/* if this is going to another job, then I have to relay
|
|
|
|
|
* it through the HNP as I have no idea how many procs
|
|
|
|
|
* are in that job
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_PROC_MY_NAME->jobid != job) {
|
|
|
|
|
if (ORTE_SUCCESS != (rc = relay_via_hnp(job, buffer, tag))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
2008-05-02 14:30:07 +00:00
|
|
|
goto CLEANUP;
|
2008-03-03 20:07:02 +00:00
|
|
|
}
|
2008-03-06 18:13:05 +00:00
|
|
|
/* if this is going to the daemon job to
|
|
|
|
|
* someplace other than the daemon cmd processor, then I need to add
|
|
|
|
|
* a command to the buffer so the recipient daemons know what to do
|
2008-03-03 20:07:02 +00:00
|
|
|
*/
|
|
|
|
|
if (ORTE_RML_TAG_DAEMON != tag) {
|
2008-03-06 18:13:05 +00:00
|
|
|
/* setup a buffer to handle the additional info */
|
|
|
|
|
buf = OBJ_NEW(opal_buffer_t);
|
|
|
|
|
/* add the proper command so the daemon's know what to do */
|
2008-03-03 20:07:02 +00:00
|
|
|
command = ORTE_DAEMON_MESSAGE_LOCAL_PROCS;
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &command, 1, ORTE_DAEMON_CMD))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &job, 1, ORTE_JOBID))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &tag, 1, ORTE_RML_TAG))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
2008-03-06 18:13:05 +00:00
|
|
|
/* copy the payload into the new buffer - this is non-destructive, so our
|
|
|
|
|
* caller is still responsible for releasing any memory in the buffer they
|
|
|
|
|
* gave to us
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.copy_payload(buf, buffer))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
/* point to correct buffer to be sent */
|
|
|
|
|
bfr = buf;
|
|
|
|
|
/* send this to the daemon tag so it gets processed correctly */
|
|
|
|
|
target = ORTE_RML_TAG_DAEMON;
|
|
|
|
|
/* set the number of targets to be the number of daemons */
|
|
|
|
|
num_targets = orte_process_info.num_procs;
|
|
|
|
|
/* send it */
|
|
|
|
|
goto SEND;
|
2008-03-03 20:07:02 +00:00
|
|
|
}
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
}
|
2008-03-06 18:13:05 +00:00
|
|
|
|
|
|
|
|
/* if I am the HNP */
|
|
|
|
|
if (orte_process_info.hnp) {
|
|
|
|
|
orte_job_t *jdata;
|
|
|
|
|
|
|
|
|
|
/* if this is going to the daemon job */
|
|
|
|
|
if (ORTE_PROC_MY_NAME->jobid == job) {
|
|
|
|
|
/* if this is going someplace other than the daemon cmd
|
|
|
|
|
* processor, then I need to add a command to the buffer
|
|
|
|
|
* so the recipient daemons know what to do
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_RML_TAG_DAEMON != tag) {
|
|
|
|
|
/* since we have to pack some additional info into the buffer
|
|
|
|
|
* for this case, we create a new buffer to contain all the
|
|
|
|
|
* info needed plus the payload
|
|
|
|
|
*/
|
|
|
|
|
buf = OBJ_NEW(opal_buffer_t);
|
|
|
|
|
/* if the target isn't the daemon tag, then we have to add the proper
|
|
|
|
|
* command so the daemon's know what to do
|
|
|
|
|
*/
|
|
|
|
|
command = ORTE_DAEMON_MESSAGE_LOCAL_PROCS;
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &command, 1, ORTE_DAEMON_CMD))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &job, 1, ORTE_JOBID))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(buf, &tag, 1, ORTE_RML_TAG))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
/* copy the payload into the new buffer - this is non-destructive, so our
|
|
|
|
|
* caller is still responsible for releasing any memory in the buffer they
|
|
|
|
|
* gave to us
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.copy_payload(buf, buffer))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
/* point to correct buffer to be sent */
|
|
|
|
|
bfr = buf;
|
|
|
|
|
/* send this to the daemon tag so it gets processed correctly */
|
|
|
|
|
target = ORTE_RML_TAG_DAEMON;
|
|
|
|
|
/* set the number of targets to be the number of daemons */
|
|
|
|
|
num_targets = orte_process_info.num_procs;
|
|
|
|
|
/* send it */
|
|
|
|
|
goto SEND;
|
|
|
|
|
} else {
|
|
|
|
|
/* if already going to the daemon tag, then just point to
|
|
|
|
|
* the right places and send it
|
|
|
|
|
*/
|
|
|
|
|
bfr = buffer;
|
|
|
|
|
target = tag;
|
|
|
|
|
num_targets = orte_process_info.num_procs;
|
|
|
|
|
goto SEND;
|
|
|
|
|
}
|
2008-03-03 20:07:02 +00:00
|
|
|
}
|
2008-03-06 18:13:05 +00:00
|
|
|
/* if this is going to any other job,
|
|
|
|
|
* then I need to know the number of procs in that job so I can
|
|
|
|
|
* send it
|
2008-03-03 20:07:02 +00:00
|
|
|
*/
|
2008-03-06 18:13:05 +00:00
|
|
|
if (NULL == (jdata = orte_get_job_data_object(job))) {
|
|
|
|
|
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
|
|
|
|
|
rc = ORTE_ERR_NOT_FOUND;
|
2008-03-03 20:07:02 +00:00
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
2008-03-06 18:13:05 +00:00
|
|
|
/* set the number of targets */
|
|
|
|
|
num_targets = jdata->num_procs;
|
|
|
|
|
/* set the tag */
|
|
|
|
|
target = tag;
|
2008-03-03 20:07:02 +00:00
|
|
|
/* point to correct buffer to be sent */
|
2008-03-06 18:13:05 +00:00
|
|
|
bfr = buffer;
|
|
|
|
|
/* send it */
|
|
|
|
|
goto SEND;
|
2008-03-03 20:07:02 +00:00
|
|
|
}
|
2008-03-06 18:13:05 +00:00
|
|
|
|
2008-03-03 20:07:02 +00:00
|
|
|
|
2008-03-06 18:13:05 +00:00
|
|
|
SEND:
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
|
|
|
|
|
"%s xcast_direct: buffer size %ld",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
(long)buffer->bytes_used));
|
2008-03-06 18:13:05 +00:00
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
peer.jobid = job;
|
2008-03-06 18:13:05 +00:00
|
|
|
for(i=0; i<num_targets; i++) {
|
2008-02-28 01:57:57 +00:00
|
|
|
peer.vpid = i;
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
|
2008-02-28 01:57:57 +00:00
|
|
|
"%s xcast_direct: %s => %s",
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
2008-02-28 01:57:57 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_NAME_PRINT(&peer)));
|
|
|
|
|
|
2008-03-03 20:07:02 +00:00
|
|
|
/* if I am the HNP, just set things up so the cmd processor gets called.
|
|
|
|
|
* We don't want to message ourselves as this can create circular logic
|
|
|
|
|
* in the RML. Instead, this macro will set a zero-time event which will
|
|
|
|
|
* cause the buffer to be processed by the cmd processor - probably will
|
|
|
|
|
* fire right away, but that's okay
|
|
|
|
|
* The macro makes a copy of the buffer, so it's okay to release it later
|
|
|
|
|
*/
|
2008-02-28 01:57:57 +00:00
|
|
|
if (peer.jobid == ORTE_PROC_MY_NAME->jobid &&
|
|
|
|
|
peer.vpid == ORTE_PROC_MY_NAME->vpid &&
|
|
|
|
|
orte_process_info.hnp) {
|
2008-03-03 20:07:02 +00:00
|
|
|
ORTE_MESSAGE_EVENT(ORTE_PROC_MY_NAME, bfr, ORTE_RML_TAG_DAEMON, orte_daemon_cmd_processor);
|
2008-02-28 01:57:57 +00:00
|
|
|
} else {
|
2008-03-03 20:07:02 +00:00
|
|
|
if (0 > (rc = orte_rml.send_buffer(&peer, bfr, target, 0))) {
|
2008-02-28 01:57:57 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
2008-03-03 20:07:02 +00:00
|
|
|
rc = ORTE_SUCCESS;
|
2008-02-28 01:57:57 +00:00
|
|
|
}
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
}
|
|
|
|
|
rc = ORTE_SUCCESS;
|
2008-02-28 01:57:57 +00:00
|
|
|
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
CLEANUP:
|
2008-03-03 20:07:02 +00:00
|
|
|
/* release buf if used */
|
|
|
|
|
if (NULL != buf) {
|
|
|
|
|
OBJ_RELEASE(buf);
|
|
|
|
|
}
|
2008-03-17 19:34:36 +00:00
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm: xcast_direct completed",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
static opal_list_t* next_recips(orte_grpcomm_mode_t mode)
|
2008-02-28 01:57:57 +00:00
|
|
|
{
|
2008-04-09 22:10:53 +00:00
|
|
|
/* check the mode to select the proper algo */
|
|
|
|
|
switch (mode) {
|
|
|
|
|
case ORTE_GRPCOMM_CHAIN:
|
|
|
|
|
ORTE_ERROR_LOG(ORTE_ERR_NOT_IMPLEMENTED);
|
|
|
|
|
return NULL;
|
|
|
|
|
break;
|
|
|
|
|
case ORTE_GRPCOMM_BINOMIAL:
|
|
|
|
|
return my_children;
|
|
|
|
|
break;
|
|
|
|
|
case ORTE_GRPCOMM_LINEAR:
|
|
|
|
|
ORTE_ERROR_LOG(ORTE_ERR_NOT_IMPLEMENTED);
|
|
|
|
|
return NULL;
|
|
|
|
|
break;
|
|
|
|
|
default:
|
|
|
|
|
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
|
|
|
|
|
return NULL;
|
|
|
|
|
break;
|
2008-02-28 01:57:57 +00:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
|
|
|
|
|
static bool barrier_recvd;
|
|
|
|
|
static bool barrier_timer;
|
|
|
|
|
|
|
|
|
|
static void barrier_recv(int status, orte_process_name_t* sender,
|
|
|
|
|
opal_buffer_t *buffer,
|
|
|
|
|
orte_rml_tag_t tag, void *cbdata)
|
|
|
|
|
{
|
|
|
|
|
/* flag as recvd */
|
|
|
|
|
barrier_recvd = true;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void barrier_timer_recv(int status, orte_process_name_t* sender,
|
|
|
|
|
opal_buffer_t *buffer,
|
|
|
|
|
orte_rml_tag_t tag, void *cbdata)
|
2008-02-28 01:57:57 +00:00
|
|
|
{
|
2008-04-09 22:10:53 +00:00
|
|
|
barrier_timer = true;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static int find_parent(int rank, int parent, int me, int num_procs,
|
|
|
|
|
int *num_children, opal_list_t *children)
|
|
|
|
|
{
|
|
|
|
|
int i, bitmap, peer, hibit, mask, found;
|
|
|
|
|
orte_namelist_t *child;
|
|
|
|
|
|
|
|
|
|
/* is this me? */
|
|
|
|
|
if (me == rank) {
|
|
|
|
|
bitmap = opal_cube_dim(num_procs);
|
|
|
|
|
|
|
|
|
|
hibit = opal_hibit(rank, bitmap);
|
|
|
|
|
--bitmap;
|
|
|
|
|
|
|
|
|
|
for (i = hibit + 1, mask = 1 << i; i <= bitmap; ++i, mask <<= 1) {
|
|
|
|
|
peer = rank | mask;
|
|
|
|
|
if (peer < num_procs) {
|
|
|
|
|
if (NULL != children) {
|
|
|
|
|
child = OBJ_NEW(orte_namelist_t);
|
|
|
|
|
child->name.jobid = ORTE_PROC_MY_NAME->jobid;
|
|
|
|
|
child->name.vpid = peer;
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((3, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic find-parent found child %s",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_NAME_PRINT(&child->name)));
|
|
|
|
|
|
|
|
|
|
opal_list_append(children, &child->item);
|
|
|
|
|
}
|
|
|
|
|
(*num_children)++;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((3, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic find-parent found parent %d",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
parent));
|
|
|
|
|
return parent;
|
|
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
/* find the children of this rank */
|
|
|
|
|
bitmap = opal_cube_dim(num_procs);
|
2008-02-28 01:57:57 +00:00
|
|
|
|
|
|
|
|
hibit = opal_hibit(rank, bitmap);
|
|
|
|
|
--bitmap;
|
|
|
|
|
|
|
|
|
|
for (i = hibit + 1, mask = 1 << i; i <= bitmap; ++i, mask <<= 1) {
|
|
|
|
|
peer = rank | mask;
|
2008-04-09 22:10:53 +00:00
|
|
|
if (peer < num_procs) {
|
|
|
|
|
/* execute compute on this child */
|
|
|
|
|
if (0 <= (found = find_parent(peer, rank, me, num_procs, num_children, children))) {
|
|
|
|
|
return found;
|
2008-02-28 01:57:57 +00:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
2008-04-09 22:10:53 +00:00
|
|
|
return -1;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
static int barrier(void)
|
|
|
|
|
{
|
|
|
|
|
opal_buffer_t buf;
|
|
|
|
|
orte_daemon_cmd_flag_t command=ORTE_DAEMON_COLL_CMD;
|
|
|
|
|
orte_grpcomm_coll_t coll_type=ORTE_GRPCOMM_BARRIER;
|
|
|
|
|
int rc;
|
|
|
|
|
struct timeval ompistart, ompistop;
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic entering barrier",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
|
|
|
|
if (orte_timing && ORTE_PROC_MY_NAME->vpid == 0) {
|
|
|
|
|
gettimeofday(&ompistart, NULL);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* everyone sends barrier to local daemon */
|
|
|
|
|
OBJ_CONSTRUCT(&buf, opal_buffer_t);
|
|
|
|
|
/* tell the daemon to collect the data */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(&buf, &command, 1, ORTE_DAEMON_CMD))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
OBJ_DESTRUCT(&buf);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
/* tell the daemon we are doing a barrier */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(&buf, &coll_type, 1, ORTE_GRPCOMM_COLL_T))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
OBJ_DESTRUCT(&buf);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
/* send to local daemon */
|
|
|
|
|
if (0 > (rc = orte_rml.send_buffer(ORTE_PROC_MY_DAEMON, &buf, ORTE_RML_TAG_DAEMON, 0))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
OBJ_DESTRUCT(&buf);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
OBJ_DESTRUCT(&buf);
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic barrier sent",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
|
|
|
|
/* now receive the release. Be sure to do this in
|
|
|
|
|
* a manner that allows us to return without being in a recv!
|
|
|
|
|
*/
|
|
|
|
|
barrier_recvd = false;
|
|
|
|
|
rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_BARRIER,
|
|
|
|
|
ORTE_RML_NON_PERSISTENT, barrier_recv, NULL);
|
|
|
|
|
if (rc != ORTE_SUCCESS) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
ORTE_PROGRESSED_WAIT(barrier_recvd, 0, 1);
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic received barrier release",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
|
|
|
|
if (orte_timing) {
|
|
|
|
|
if (ORTE_PROC_MY_NAME->vpid == 0) {
|
|
|
|
|
/* setup a receive to hear when the rank=N proc has received the data
|
|
|
|
|
* release - in most xcast schemes, this will always be the final recvr
|
|
|
|
|
*/
|
|
|
|
|
barrier_timer = false;
|
|
|
|
|
orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_COLLECTIVE_TIMER,
|
|
|
|
|
ORTE_RML_NON_PERSISTENT, barrier_timer_recv, NULL);
|
|
|
|
|
if (rc != ORTE_SUCCESS) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
ORTE_PROGRESSED_WAIT(barrier_timer, 0, 1);
|
|
|
|
|
gettimeofday(&ompistop, NULL);
|
|
|
|
|
opal_output(0, "%s time to complete barrier %ld usec",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
(long int)((ompistop.tv_sec - ompistart.tv_sec)*1000000 +
|
|
|
|
|
(ompistop.tv_usec - ompistart.tv_usec)));
|
|
|
|
|
} else if (ORTE_PROC_MY_NAME->vpid == orte_process_info.num_procs-1) {
|
|
|
|
|
/* if we are rank=N, send a message back to indicate
|
|
|
|
|
* the xcast completed for timing purposes
|
|
|
|
|
*/
|
|
|
|
|
orte_process_name_t name;
|
|
|
|
|
|
|
|
|
|
name.jobid = ORTE_PROC_MY_NAME->jobid;
|
|
|
|
|
name.vpid = 0;
|
|
|
|
|
OBJ_CONSTRUCT(&buf, opal_buffer_t);
|
|
|
|
|
if (0 > (rc = orte_rml.send_buffer(&name,&buf,ORTE_RML_TAG_COLLECTIVE_TIMER,0))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
OBJ_DESTRUCT(&buf);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
rc = ORTE_SUCCESS;
|
|
|
|
|
OBJ_DESTRUCT(&buf);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
static opal_buffer_t *allgather_buf;
|
|
|
|
|
static orte_std_cntr_t allgather_complete;
|
|
|
|
|
|
|
|
|
|
static void allgather_recv(int status, orte_process_name_t* sender,
|
|
|
|
|
opal_buffer_t *buffer,
|
|
|
|
|
orte_rml_tag_t tag, void *cbdata)
|
2008-03-03 20:07:02 +00:00
|
|
|
{
|
2008-04-09 22:10:53 +00:00
|
|
|
int rc;
|
2008-03-03 20:07:02 +00:00
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
/* xfer the data */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.copy_payload(allgather_buf, buffer))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
|
|
|
|
allgather_complete = true;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static int allgather(opal_buffer_t *sbuf, opal_buffer_t *rbuf)
|
|
|
|
|
{
|
|
|
|
|
int rc;
|
|
|
|
|
orte_daemon_cmd_flag_t command=ORTE_DAEMON_COLL_CMD;
|
|
|
|
|
struct timeval ompistart, ompistop;
|
|
|
|
|
opal_buffer_t coll;
|
|
|
|
|
orte_grpcomm_coll_t coll_type=ORTE_GRPCOMM_ALLGATHER;
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic entering allgather",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
|
|
|
|
if (orte_timing && ORTE_PROC_MY_NAME->vpid == 0) {
|
|
|
|
|
gettimeofday(&ompistart, NULL);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* everyone sends data to their local daemon */
|
|
|
|
|
OBJ_CONSTRUCT(&coll, opal_buffer_t);
|
|
|
|
|
/* tell the daemon to collect the data */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(&coll, &command, 1, ORTE_DAEMON_CMD))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
OBJ_DESTRUCT(&coll);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
/* tell the daemon we are doing an allgather */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(&coll, &coll_type, 1, ORTE_GRPCOMM_COLL_T))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
OBJ_DESTRUCT(&coll);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
/* add our data to it */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.copy_payload(&coll, sbuf))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
OBJ_DESTRUCT(&coll);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
/* send to local daemon */
|
|
|
|
|
if (0 > (rc = orte_rml.send_buffer(ORTE_PROC_MY_DAEMON, &coll, ORTE_RML_TAG_DAEMON, 0))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
OBJ_DESTRUCT(&coll);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
OBJ_DESTRUCT(&coll);
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic allgather buffer sent",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
|
|
|
|
/* setup the buffer that will recv the results */
|
|
|
|
|
allgather_buf = OBJ_NEW(opal_buffer_t);
|
|
|
|
|
|
|
|
|
|
/* now receive the final result. Be sure to do this in
|
|
|
|
|
* a manner that allows us to return without being in a recv!
|
|
|
|
|
*/
|
|
|
|
|
allgather_complete = false;
|
|
|
|
|
rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_ALLGATHER,
|
|
|
|
|
ORTE_RML_NON_PERSISTENT, allgather_recv, NULL);
|
|
|
|
|
if (rc != ORTE_SUCCESS) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
2008-03-03 20:07:02 +00:00
|
|
|
}
|
|
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
ORTE_PROGRESSED_WAIT(allgather_complete, 0, 1);
|
|
|
|
|
|
|
|
|
|
/* copy payload to the caller's buffer */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.copy_payload(rbuf, allgather_buf))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
OBJ_RELEASE(allgather_buf);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
OBJ_RELEASE(allgather_buf);
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
|
|
|
|
|
"%s allgather buffer received",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
|
|
|
|
if (orte_timing) {
|
|
|
|
|
if (ORTE_PROC_MY_NAME->vpid == 0) {
|
|
|
|
|
/* setup a receive to hear when the rank=N proc has received the data
|
|
|
|
|
* release - in most xcast schemes, this will always be the final recvr
|
|
|
|
|
*/
|
|
|
|
|
barrier_timer = false;
|
|
|
|
|
rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_COLLECTIVE_TIMER,
|
|
|
|
|
ORTE_RML_NON_PERSISTENT, barrier_timer_recv, NULL);
|
|
|
|
|
if (ORTE_SUCCESS != rc) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
ORTE_PROGRESSED_WAIT(barrier_timer, 0, 1);
|
|
|
|
|
gettimeofday(&ompistop, NULL);
|
|
|
|
|
opal_output(0, "%s allgather: time to complete %ld usec",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
(long int)((ompistop.tv_sec - ompistart.tv_sec)*1000000 +
|
|
|
|
|
(ompistop.tv_usec - ompistart.tv_usec)));
|
|
|
|
|
} else if (ORTE_PROC_MY_NAME->vpid == orte_process_info.num_procs-1) {
|
|
|
|
|
/* if we are rank=N, send a message back to indicate
|
|
|
|
|
* the xcast completed for timing purposes
|
|
|
|
|
*/
|
|
|
|
|
orte_process_name_t name;
|
|
|
|
|
opal_buffer_t buf;
|
|
|
|
|
|
|
|
|
|
name.jobid = ORTE_PROC_MY_NAME->jobid;
|
|
|
|
|
name.vpid = 0;
|
|
|
|
|
OBJ_CONSTRUCT(&buf, opal_buffer_t);
|
|
|
|
|
if (0 > (rc = orte_rml.send_buffer(&name,&buf,ORTE_RML_TAG_COLLECTIVE_TIMER,0))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
rc = ORTE_SUCCESS;
|
|
|
|
|
OBJ_DESTRUCT(&buf);
|
2008-03-03 20:07:02 +00:00
|
|
|
}
|
|
|
|
|
}
|
2008-04-09 22:10:53 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic allgather completed",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
2008-03-03 20:07:02 +00:00
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
static orte_std_cntr_t collective_num_recvd;
|
|
|
|
|
static bool collective_failed;
|
|
|
|
|
static opal_buffer_t *collection;
|
|
|
|
|
static orte_std_cntr_t num_contributors;
|
|
|
|
|
|
|
|
|
|
static void collective_recv(int status, orte_process_name_t* sender,
|
|
|
|
|
opal_buffer_t *buffer,
|
|
|
|
|
orte_rml_tag_t tag, void *cbdata)
|
2008-02-28 01:57:57 +00:00
|
|
|
{
|
|
|
|
|
int rc;
|
2008-04-09 22:10:53 +00:00
|
|
|
orte_std_cntr_t contributors, cnt;
|
2008-02-28 01:57:57 +00:00
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
/* extract the #contributors */
|
|
|
|
|
cnt=1;
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.unpack(buffer, &contributors, &cnt, ORTE_STD_CNTR))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
|
|
|
|
num_contributors += contributors;
|
|
|
|
|
|
|
|
|
|
/* xfer the data */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.copy_payload(collection, buffer))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
collective_failed = true;
|
|
|
|
|
}
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic collective recv - got %d bytes from %s with %d contributors",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
(int)(buffer->bytes_used-sizeof(orte_std_cntr_t)),
|
|
|
|
|
ORTE_NAME_PRINT(sender), (int)contributors));
|
|
|
|
|
|
|
|
|
|
/* reissue the recv */
|
|
|
|
|
rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_DAEMON_COLLECTIVE,
|
|
|
|
|
ORTE_RML_NON_PERSISTENT, collective_recv, NULL);
|
|
|
|
|
if (rc != ORTE_SUCCESS) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
collective_failed = true;
|
|
|
|
|
}
|
2008-04-23 15:43:44 +00:00
|
|
|
/* bump counter */
|
|
|
|
|
++collective_num_recvd;
|
2008-04-09 22:10:53 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
static int daemon_leader(orte_jobid_t jobid,
|
|
|
|
|
orte_std_cntr_t num_local_contributors,
|
|
|
|
|
orte_grpcomm_coll_t type,
|
|
|
|
|
opal_buffer_t *data,
|
2008-04-30 19:49:53 +00:00
|
|
|
bool hnp_has_local_procs)
|
2008-04-09 22:10:53 +00:00
|
|
|
{
|
|
|
|
|
int rc;
|
|
|
|
|
opal_buffer_t buf;
|
|
|
|
|
int num_children;
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic daemon_collective - I am the leader!",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
2008-04-30 19:49:53 +00:00
|
|
|
if (hnp_has_local_procs) {
|
2008-04-09 22:10:53 +00:00
|
|
|
/* if everyone is participating, then I must be the HNP,
|
|
|
|
|
* so the #children is just the #children determined for
|
|
|
|
|
* my outgoing xcast
|
|
|
|
|
*/
|
|
|
|
|
num_children = my_num_children;
|
2008-04-30 19:49:53 +00:00
|
|
|
} else {
|
2008-04-09 22:10:53 +00:00
|
|
|
/* if the HNP has no local procs, then it won't
|
|
|
|
|
* know that a collective is underway, so that means
|
|
|
|
|
* I must be rank=1. The number of messages I must get
|
|
|
|
|
* therefore consists of both my children + all other
|
|
|
|
|
* children of rank=0 as they will redirect their messages
|
|
|
|
|
* to me
|
|
|
|
|
*/
|
|
|
|
|
num_children = 0;
|
|
|
|
|
/* find #children for rank=0 */
|
|
|
|
|
find_parent(0, 0, 0, orte_process_info.num_procs, &num_children, NULL);
|
|
|
|
|
/* I am one of those children, so we should get num_children-1 of
|
|
|
|
|
* my peers sending to me, plus my own children
|
|
|
|
|
*/
|
|
|
|
|
num_children = num_children - 1 + my_num_children;
|
|
|
|
|
}
|
2008-04-30 19:49:53 +00:00
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
/* setup to recv the messages from my children */
|
|
|
|
|
collective_num_recvd = 0;
|
|
|
|
|
collective_failed = false;
|
|
|
|
|
collection = OBJ_NEW(opal_buffer_t);
|
|
|
|
|
num_contributors = num_local_contributors; /* seed with the number I added */
|
|
|
|
|
|
|
|
|
|
/* ensure my data gets included in the outcome */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.copy_payload(collection, data))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
OBJ_RELEASE(collection);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* if we have children, get their messages */
|
|
|
|
|
if (0 < num_children) {
|
|
|
|
|
/* post the non-blocking recv */
|
|
|
|
|
rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_DAEMON_COLLECTIVE,
|
|
|
|
|
ORTE_RML_NON_PERSISTENT, collective_recv, NULL);
|
|
|
|
|
if (rc != ORTE_SUCCESS) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
OBJ_RELEASE(collection);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
ORTE_PROGRESSED_WAIT(collective_failed, collective_num_recvd, num_children);
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic daemon_collective - leader has received collective from %d children",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), num_children));
|
|
|
|
|
|
|
|
|
|
/* cancel the lingering recv */
|
2008-04-23 15:43:44 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml.recv_cancel(ORTE_NAME_WILDCARD, ORTE_RML_TAG_DAEMON_COLLECTIVE)) &&
|
|
|
|
|
ORTE_ERR_NOT_FOUND != rc) {
|
2008-04-09 22:10:53 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
OBJ_RELEASE(collection);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
OBJ_CONSTRUCT(&buf, opal_buffer_t);
|
|
|
|
|
|
|
|
|
|
if (ORTE_GRPCOMM_BARRIER == type) {
|
|
|
|
|
if (ORTE_SUCCESS != (rc = xcast(jobid, &buf, ORTE_RML_TAG_BARRIER))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
|
|
|
|
} else if (ORTE_GRPCOMM_ALLGATHER == type) {
|
|
|
|
|
/* send the data */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(&buf, &num_contributors, 1, ORTE_STD_CNTR))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto cleanup;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.copy_payload(&buf, collection))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto cleanup;
|
|
|
|
|
}
|
|
|
|
|
if (ORTE_SUCCESS != (rc = xcast(jobid, &buf, ORTE_RML_TAG_ALLGATHER))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
/* no other collectives currently supported! */
|
|
|
|
|
ORTE_ERROR_LOG(ORTE_ERR_NOT_IMPLEMENTED);
|
|
|
|
|
rc = ORTE_ERR_NOT_IMPLEMENTED;
|
2008-02-28 01:57:57 +00:00
|
|
|
}
|
2008-04-09 22:10:53 +00:00
|
|
|
|
|
|
|
|
cleanup:
|
|
|
|
|
OBJ_RELEASE(collection);
|
|
|
|
|
OBJ_DESTRUCT(&buf);
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic daemon_collective - leader has completed collective",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
static int daemon_collective(orte_jobid_t jobid,
|
|
|
|
|
orte_std_cntr_t num_local_contributors,
|
|
|
|
|
orte_grpcomm_coll_t type,
|
|
|
|
|
opal_buffer_t *data,
|
2008-04-30 19:49:53 +00:00
|
|
|
bool hnp_has_local_procs)
|
2008-04-09 22:10:53 +00:00
|
|
|
{
|
|
|
|
|
orte_process_name_t lead, parent;
|
|
|
|
|
int num_children;
|
|
|
|
|
opal_buffer_t buf;
|
|
|
|
|
int rc;
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
2008-04-30 19:49:53 +00:00
|
|
|
"%s grpcomm:basic daemon_collective entered - %s",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
hnp_has_local_procs ? "HNP HAS LOCAL PROCS" : "HNP DOES NOT HAVE LOCAL PROCS"));
|
2008-04-09 22:10:53 +00:00
|
|
|
|
|
|
|
|
parent.jobid = ORTE_PROC_MY_NAME->jobid;
|
|
|
|
|
lead.jobid = ORTE_PROC_MY_NAME->jobid;
|
|
|
|
|
|
|
|
|
|
/* if the participation is full, then the HNP is the lead */
|
2008-04-30 19:49:53 +00:00
|
|
|
if (hnp_has_local_procs) {
|
2008-04-09 22:10:53 +00:00
|
|
|
lead.vpid = ORTE_PROC_MY_HNP->vpid;
|
2008-04-30 19:49:53 +00:00
|
|
|
} else {
|
2008-04-09 22:10:53 +00:00
|
|
|
/* if the HNP has no local procs, then it won't
|
|
|
|
|
* know that a collective is underway, so let
|
|
|
|
|
* rank=1 be the lead
|
|
|
|
|
*/
|
|
|
|
|
lead.vpid = 1;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* if I am the lead, do my own thing */
|
|
|
|
|
if (ORTE_PROC_MY_NAME->vpid == lead.vpid) {
|
2008-04-30 19:49:53 +00:00
|
|
|
return daemon_leader(jobid, num_local_contributors, type, data, hnp_has_local_procs);
|
2008-04-09 22:10:53 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/* I am NOT the lead, so I first must figure out how many children
|
|
|
|
|
* I need to collect messages from and who my parent will be
|
|
|
|
|
*/
|
|
|
|
|
|
2008-04-30 19:49:53 +00:00
|
|
|
if (hnp_has_local_procs) {
|
2008-04-09 22:10:53 +00:00
|
|
|
/* everyone is participating, so my parent and
|
|
|
|
|
* num_children can be as initially computed
|
|
|
|
|
*/
|
|
|
|
|
parent.vpid = my_parent.vpid;
|
|
|
|
|
num_children = my_num_children;
|
2008-04-30 19:49:53 +00:00
|
|
|
} else {
|
2008-04-09 22:10:53 +00:00
|
|
|
/* if the HNP has no local procs, then it won't
|
|
|
|
|
* know that a collective is underway, so we need
|
|
|
|
|
* to send to rank=1 if our parent would have been
|
|
|
|
|
* rank=0. Our num_children, though,
|
|
|
|
|
* remains unchanged
|
|
|
|
|
*/
|
|
|
|
|
if (0 == my_parent.vpid) {
|
|
|
|
|
parent.vpid = 1;
|
|
|
|
|
} else {
|
|
|
|
|
/* just send as normal */
|
|
|
|
|
parent.vpid = my_parent.vpid;
|
|
|
|
|
}
|
|
|
|
|
num_children = my_num_children;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic daemon_collective preparing to receive from %d children",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
num_children));
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
/* setup for collecting data */
|
|
|
|
|
collection = OBJ_NEW(opal_buffer_t);
|
|
|
|
|
num_contributors = num_local_contributors; /* seed with the number I added */
|
|
|
|
|
|
|
|
|
|
/* ensure my data gets included in the outcome */
|
|
|
|
|
opal_dss.copy_payload(collection, data);
|
|
|
|
|
|
|
|
|
|
/* if num_children > 0, setup recv's to wait until we hear from
|
|
|
|
|
* them all - the recv will look just like that for the leader,
|
|
|
|
|
* collecting data and #contributors
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
if (0 < num_children) {
|
|
|
|
|
/* setup to recv the messages from my children */
|
|
|
|
|
collective_num_recvd = 0;
|
|
|
|
|
collective_failed = false;
|
|
|
|
|
|
|
|
|
|
/* post the non-blocking recv */
|
|
|
|
|
rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_DAEMON_COLLECTIVE,
|
|
|
|
|
ORTE_RML_NON_PERSISTENT, collective_recv, NULL);
|
|
|
|
|
if (rc != ORTE_SUCCESS) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
ORTE_PROGRESSED_WAIT(collective_failed, collective_num_recvd, num_children);
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic daemon_collective - I have received collective from children",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
|
|
|
|
/* cancel the lingering recv */
|
2008-04-23 15:43:44 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml.recv_cancel(ORTE_NAME_WILDCARD, ORTE_RML_TAG_DAEMON_COLLECTIVE)) &&
|
|
|
|
|
ORTE_ERR_NOT_FOUND != rc) {
|
2008-04-09 22:10:53 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* construct and send message to our parent */
|
|
|
|
|
OBJ_CONSTRUCT(&buf, opal_buffer_t);
|
|
|
|
|
/* insert #contributors */
|
|
|
|
|
opal_dss.pack(&buf, &num_contributors, 1, ORTE_STD_CNTR);
|
|
|
|
|
|
|
|
|
|
if (ORTE_GRPCOMM_BARRIER == type) {
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic daemon_collective sending barrier to %s",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_NAME_PRINT(&parent)));
|
|
|
|
|
|
|
|
|
|
if (0 > (rc = orte_rml.send_buffer(&parent,&buf,ORTE_RML_TAG_DAEMON_COLLECTIVE,0))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
rc= ORTE_SUCCESS;
|
|
|
|
|
} else if (ORTE_GRPCOMM_ALLGATHER == type) {
|
|
|
|
|
/* xfer the data */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.copy_payload(&buf, collection))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
/* send the data */
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic daemon_collective sending allgather data to %s",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_NAME_PRINT(&parent)));
|
|
|
|
|
|
|
|
|
|
if (0 > (rc = orte_rml.send_buffer(&parent,&buf,ORTE_RML_TAG_DAEMON_COLLECTIVE,0))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
rc = ORTE_SUCCESS;
|
|
|
|
|
} else {
|
|
|
|
|
/* we don't currently support any other collectives */
|
|
|
|
|
ORTE_ERROR_LOG(ORTE_ERR_NOT_IMPLEMENTED);
|
|
|
|
|
rc = ORTE_ERR_NOT_IMPLEMENTED;
|
|
|
|
|
}
|
|
|
|
|
OBJ_DESTRUCT(&buf);
|
|
|
|
|
OBJ_RELEASE(collection);
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base_output,
|
|
|
|
|
"%s grpcomm:basic daemon_collective completed",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|