2007-07-20 01:34:02 +00:00
|
|
|
/*
|
|
|
|
|
* Copyright (c) 2007 Los Alamos National Security, LLC.
|
|
|
|
|
* All rights reserved.
|
|
|
|
|
* $COPYRIGHT$
|
|
|
|
|
*
|
|
|
|
|
* Additional copyrights may follow
|
|
|
|
|
*
|
|
|
|
|
* $HEADER$
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
#include "orte_config.h"
|
2008-02-28 01:57:57 +00:00
|
|
|
#include "orte/constants.h"
|
2007-07-20 01:34:02 +00:00
|
|
|
|
2010-07-20 04:22:45 +00:00
|
|
|
#include <stddef.h>
|
|
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
#include "opal/threads/condition.h"
|
|
|
|
|
#include "opal/dss/dss.h"
|
2009-03-03 22:36:58 +00:00
|
|
|
#include "opal/class/opal_bitmap.h"
|
2008-03-05 22:44:35 +00:00
|
|
|
#include "opal/class/opal_hash_table.h"
|
2009-02-14 02:26:12 +00:00
|
|
|
#include "opal/util/output.h"
|
2010-05-17 23:08:56 +00:00
|
|
|
#include "opal/util/opal_sos.h"
|
2008-03-05 22:44:35 +00:00
|
|
|
|
2007-07-20 01:34:02 +00:00
|
|
|
#include "orte/mca/errmgr/errmgr.h"
|
2008-10-31 21:10:00 +00:00
|
|
|
#include "orte/mca/ess/ess.h"
|
2008-02-28 01:57:57 +00:00
|
|
|
#include "orte/mca/rml/rml.h"
|
2009-02-14 02:26:12 +00:00
|
|
|
#include "orte/mca/rml/rml_types.h"
|
2008-02-28 01:57:57 +00:00
|
|
|
#include "orte/util/name_fns.h"
|
|
|
|
|
#include "orte/runtime/orte_globals.h"
|
2008-02-28 19:58:32 +00:00
|
|
|
#include "orte/runtime/orte_wait.h"
|
2008-06-18 03:15:56 +00:00
|
|
|
#include "orte/runtime/runtime.h"
|
2007-07-20 01:34:02 +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/rml/base/rml_contact.h"
|
|
|
|
|
|
|
|
|
|
#include "orte/mca/routed/base/base.h"
|
2008-05-05 02:59:09 +00:00
|
|
|
#include "routed_linear.h"
|
2007-07-20 01:34:02 +00:00
|
|
|
|
2008-03-26 01:00:24 +00:00
|
|
|
static int init(void);
|
|
|
|
|
static int finalize(void);
|
2008-07-21 21:37:09 +00:00
|
|
|
static int delete_route(orte_process_name_t *proc);
|
2008-03-26 01:00:24 +00:00
|
|
|
static int update_route(orte_process_name_t *target,
|
|
|
|
|
orte_process_name_t *route);
|
|
|
|
|
static orte_process_name_t get_route(orte_process_name_t *target);
|
|
|
|
|
static int init_routes(orte_jobid_t job, opal_buffer_t *ndat);
|
|
|
|
|
static int route_lost(const orte_process_name_t *route);
|
2008-05-05 02:59:09 +00:00
|
|
|
static bool route_is_defined(const orte_process_name_t *target);
|
|
|
|
|
static int update_routing_tree(void);
|
2009-01-27 19:13:56 +00:00
|
|
|
static orte_vpid_t get_routing_tree(opal_list_t *children);
|
2008-10-02 20:08:27 +00:00
|
|
|
static int get_wireup_info(opal_buffer_t *buf);
|
2008-12-20 23:23:11 +00:00
|
|
|
static int set_lifeline(orte_process_name_t *proc);
|
2010-07-17 21:03:27 +00:00
|
|
|
static size_t num_routes(void);
|
2008-03-26 01:00:24 +00:00
|
|
|
|
2010-03-12 23:57:50 +00:00
|
|
|
#if OPAL_ENABLE_FT_CR == 1
|
2008-05-05 02:59:09 +00:00
|
|
|
static int linear_ft_event(int state);
|
2008-04-23 00:17:12 +00:00
|
|
|
#endif
|
|
|
|
|
|
2008-05-05 02:59:09 +00:00
|
|
|
orte_routed_module_t orte_routed_linear_module = {
|
2008-03-26 01:00:24 +00:00
|
|
|
init,
|
|
|
|
|
finalize,
|
2008-07-21 21:37:09 +00:00
|
|
|
delete_route,
|
2008-03-26 01:00:24 +00:00
|
|
|
update_route,
|
|
|
|
|
get_route,
|
|
|
|
|
init_routes,
|
|
|
|
|
route_lost,
|
2008-05-05 02:59:09 +00:00
|
|
|
route_is_defined,
|
2008-12-20 23:23:11 +00:00
|
|
|
set_lifeline,
|
2008-05-05 02:59:09 +00:00
|
|
|
update_routing_tree,
|
|
|
|
|
get_routing_tree,
|
2008-04-23 00:17:12 +00:00
|
|
|
get_wireup_info,
|
2010-07-17 21:03:27 +00:00
|
|
|
num_routes,
|
2010-03-12 23:57:50 +00:00
|
|
|
#if OPAL_ENABLE_FT_CR == 1
|
2008-05-05 02:59:09 +00:00
|
|
|
linear_ft_event
|
2008-04-23 00:17:12 +00:00
|
|
|
#else
|
|
|
|
|
NULL
|
|
|
|
|
#endif
|
2008-03-26 01:00:24 +00:00
|
|
|
};
|
|
|
|
|
|
|
|
|
|
/* local globals */
|
|
|
|
|
static opal_condition_t cond;
|
|
|
|
|
static opal_mutex_t lock;
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
static orte_process_name_t *lifeline=NULL;
|
2008-12-20 23:23:11 +00:00
|
|
|
static orte_process_name_t local_lifeline;
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
static bool ack_recvd;
|
|
|
|
|
|
2008-03-26 01:00:24 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
static int init(void)
|
|
|
|
|
{
|
|
|
|
|
/* setup the global condition and lock */
|
|
|
|
|
OBJ_CONSTRUCT(&cond, opal_condition_t);
|
|
|
|
|
OBJ_CONSTRUCT(&lock, opal_mutex_t);
|
2008-04-23 13:01:23 +00:00
|
|
|
|
|
|
|
|
lifeline = NULL;
|
|
|
|
|
|
2008-03-26 01:00:24 +00:00
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static int finalize(void)
|
|
|
|
|
{
|
2010-07-20 08:30:47 +00:00
|
|
|
int rc;
|
2008-03-26 01:00:24 +00:00
|
|
|
|
|
|
|
|
/* if I am an application process, indicate that I am
|
|
|
|
|
* truly finalizing prior to departure
|
|
|
|
|
*/
|
2009-05-04 11:07:40 +00:00
|
|
|
if (!ORTE_PROC_IS_HNP &&
|
|
|
|
|
!ORTE_PROC_IS_DAEMON &&
|
|
|
|
|
!ORTE_PROC_IS_TOOL) {
|
2008-04-30 19:49:53 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_routed_base_register_sync(false))) {
|
2008-03-26 01:00:24 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* destruct the global condition and lock */
|
|
|
|
|
OBJ_DESTRUCT(&cond);
|
|
|
|
|
OBJ_DESTRUCT(&lock);
|
2008-04-23 13:01:23 +00:00
|
|
|
|
|
|
|
|
lifeline = NULL;
|
|
|
|
|
|
2008-03-26 01:00:24 +00:00
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
2008-07-21 21:37:09 +00:00
|
|
|
static int delete_route(orte_process_name_t *proc)
|
|
|
|
|
{
|
2010-07-20 04:47:31 +00:00
|
|
|
int i;
|
|
|
|
|
orte_routed_jobfam_t *jfam;
|
|
|
|
|
uint16_t jfamily;
|
2008-07-21 21:37:09 +00:00
|
|
|
|
|
|
|
|
if (proc->jobid == ORTE_JOBID_INVALID ||
|
|
|
|
|
proc->vpid == ORTE_VPID_INVALID) {
|
|
|
|
|
return ORTE_ERR_BAD_PARAM;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* if I am an application process, I don't have any routes
|
|
|
|
|
* so there is nothing for me to do
|
|
|
|
|
*/
|
2009-05-04 11:07:40 +00:00
|
|
|
if (!ORTE_PROC_IS_HNP && !ORTE_PROC_IS_DAEMON &&
|
|
|
|
|
!ORTE_PROC_IS_TOOL) {
|
2008-07-21 21:37:09 +00:00
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
|
|
|
|
"%s routed_binomial_delete_route for %s",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_NAME_PRINT(proc)));
|
2008-07-21 21:37:09 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
/* if this is from a different job family, then I need to
|
|
|
|
|
* look it up appropriately
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_JOB_FAMILY(proc->jobid) != ORTE_JOB_FAMILY(ORTE_PROC_MY_NAME->jobid)) {
|
|
|
|
|
|
|
|
|
|
/* if I am a daemon, then I will automatically route
|
|
|
|
|
* anything to this job family via my HNP - so I have nothing
|
|
|
|
|
* in my routing table and thus have nothing to do
|
|
|
|
|
* here, just return
|
|
|
|
|
*/
|
2009-05-04 11:07:40 +00:00
|
|
|
if (ORTE_PROC_IS_DAEMON) {
|
2008-07-21 21:37:09 +00:00
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
2010-07-20 04:47:31 +00:00
|
|
|
/* see if this job family is present */
|
|
|
|
|
jfamily = ORTE_JOB_FAMILY(proc->jobid);
|
2010-07-20 08:30:47 +00:00
|
|
|
for (i=0; i < orte_routed_jobfams.size; i++) {
|
|
|
|
|
if (NULL == (jfam = (orte_routed_jobfam_t*)opal_pointer_array_get_item(&orte_routed_jobfams, i))) {
|
2010-07-20 04:47:31 +00:00
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
if (jfam->job_family == jfamily) {
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_routed_base_output,
|
|
|
|
|
"%s routed_binomial: deleting route to %s",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_JOB_FAMILY_PRINT(proc->jobid)));
|
2010-07-20 08:30:47 +00:00
|
|
|
opal_pointer_array_set_item(&orte_routed_jobfams, i, NULL);
|
2010-07-20 04:47:31 +00:00
|
|
|
OBJ_RELEASE(jfam);
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
2008-07-21 21:37:09 +00:00
|
|
|
}
|
|
|
|
|
/* not present - nothing to do */
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
2008-10-31 21:10:00 +00:00
|
|
|
/* THIS CAME FROM OUR OWN JOB FAMILY...there is nothing
|
|
|
|
|
* to do here. The routes will be redefined when we update
|
|
|
|
|
* the routing tree
|
2008-07-21 21:37:09 +00:00
|
|
|
*/
|
2008-10-31 21:10:00 +00:00
|
|
|
|
|
|
|
|
return ORTE_SUCCESS;
|
2008-07-21 21:37:09 +00:00
|
|
|
}
|
|
|
|
|
|
2008-03-26 01:00:24 +00:00
|
|
|
static int update_route(orte_process_name_t *target,
|
|
|
|
|
orte_process_name_t *route)
|
2007-07-20 01:34:02 +00:00
|
|
|
{
|
2010-07-20 04:47:31 +00:00
|
|
|
int i;
|
|
|
|
|
orte_routed_jobfam_t *jfam;
|
|
|
|
|
uint16_t jfamily;
|
2008-02-28 01:57:57 +00:00
|
|
|
|
2007-07-25 03:55:34 +00:00
|
|
|
if (target->jobid == ORTE_JOBID_INVALID ||
|
2007-07-20 01:34:02 +00:00
|
|
|
target->vpid == ORTE_VPID_INVALID) {
|
|
|
|
|
return ORTE_ERR_BAD_PARAM;
|
|
|
|
|
}
|
|
|
|
|
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
/* if I am an application process, we don't update the route since
|
|
|
|
|
* we automatically route everything through the local daemon
|
|
|
|
|
*/
|
2009-05-11 14:11:44 +00:00
|
|
|
if (ORTE_PROC_IS_APP) {
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
2009-02-06 15:41:53 +00:00
|
|
|
/* if the job family is zero, then this is going to a local slave,
|
|
|
|
|
* so the path is direct and there is nothing to do here
|
|
|
|
|
*/
|
|
|
|
|
if (0 == ORTE_JOB_FAMILY(target->jobid)) {
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
2008-06-09 14:53:58 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
2008-05-05 02:59:09 +00:00
|
|
|
"%s routed_linear_update: %s --> %s",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_NAME_PRINT(target),
|
|
|
|
|
ORTE_NAME_PRINT(route)));
|
2007-07-20 01:34:02 +00:00
|
|
|
|
|
|
|
|
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
/* if this is from a different job family, then I need to
|
|
|
|
|
* track how to send messages to it
|
2008-02-28 01:57:57 +00:00
|
|
|
*/
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
if (ORTE_JOB_FAMILY(target->jobid) != ORTE_JOB_FAMILY(ORTE_PROC_MY_NAME->jobid)) {
|
|
|
|
|
|
|
|
|
|
/* if I am a daemon, then I will automatically route
|
|
|
|
|
* anything to this job family via my HNP - so nothing to do
|
|
|
|
|
* here, just return
|
2008-02-28 01:57:57 +00:00
|
|
|
*/
|
2009-05-04 11:07:40 +00:00
|
|
|
if (ORTE_PROC_IS_DAEMON) {
|
2008-02-28 01:57:57 +00:00
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
|
|
|
|
"%s routed_linear_update: diff job family routing job %s --> %s",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_JOBID_PRINT(target->jobid),
|
|
|
|
|
ORTE_NAME_PRINT(route)));
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
|
2010-07-20 04:47:31 +00:00
|
|
|
/* see if this target is already present */
|
|
|
|
|
jfamily = ORTE_JOB_FAMILY(target->jobid);
|
2010-07-20 08:30:47 +00:00
|
|
|
for (i=0; i < orte_routed_jobfams.size; i++) {
|
|
|
|
|
if (NULL == (jfam = (orte_routed_jobfam_t*)opal_pointer_array_get_item(&orte_routed_jobfams, i))) {
|
2010-07-20 04:47:31 +00:00
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
if (jfam->job_family == jfamily) {
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_routed_base_output,
|
|
|
|
|
"%s routed_linear: updating route to %s via %s",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_JOB_FAMILY_PRINT(target->jobid),
|
|
|
|
|
ORTE_NAME_PRINT(route)));
|
|
|
|
|
jfam->route.jobid = route->jobid;
|
|
|
|
|
jfam->route.vpid = route->vpid;
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
}
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
|
|
|
|
|
/* not there, so add the route FOR THE JOB FAMILY*/
|
2010-07-20 04:47:31 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_routed_base_output,
|
|
|
|
|
"%s routed_linear: adding route to %s",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_JOB_FAMILY_PRINT(target->jobid)));
|
|
|
|
|
jfam = OBJ_NEW(orte_routed_jobfam_t);
|
|
|
|
|
jfam->job_family = jfamily;
|
|
|
|
|
jfam->route.jobid = route->jobid;
|
|
|
|
|
jfam->route.vpid = route->vpid;
|
2010-07-20 08:30:47 +00:00
|
|
|
opal_pointer_array_add(&orte_routed_jobfams, jfam);
|
2010-07-20 04:47:31 +00:00
|
|
|
return ORTE_SUCCESS;
|
2008-02-28 01:57:57 +00:00
|
|
|
}
|
|
|
|
|
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
/* THIS CAME FROM OUR OWN JOB FAMILY... */
|
|
|
|
|
|
2009-03-05 21:50:47 +00:00
|
|
|
opal_output(0, "%s CALL TO UPDATE ROUTE FOR OWN JOB FAMILY", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
|
2008-07-21 21:37:09 +00:00
|
|
|
|
2007-07-20 01:34:02 +00:00
|
|
|
return ORTE_ERR_NOT_SUPPORTED;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2008-03-26 01:00:24 +00:00
|
|
|
static orte_process_name_t get_route(orte_process_name_t *target)
|
2007-07-20 01:34:02 +00:00
|
|
|
{
|
2008-10-31 21:10:00 +00:00
|
|
|
orte_process_name_t *ret, daemon;
|
2010-07-20 04:47:31 +00:00
|
|
|
int i;
|
|
|
|
|
orte_routed_jobfam_t *jfam;
|
|
|
|
|
uint16_t jfamily;
|
2007-07-20 01:34:02 +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
|
|
|
/* if it is me, then the route is just direct */
|
2008-02-28 01:57:57 +00:00
|
|
|
if (OPAL_EQUAL == opal_dss.compare(ORTE_PROC_MY_NAME, target, ORTE_NAME)) {
|
2008-03-05 22:44:35 +00:00
|
|
|
ret = target;
|
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
|
|
|
goto found;
|
|
|
|
|
}
|
|
|
|
|
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
/* if I am an application process, always route via my local daemon */
|
2009-05-11 14:11:44 +00:00
|
|
|
if (ORTE_PROC_IS_APP) {
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
ret = ORTE_PROC_MY_DAEMON;
|
|
|
|
|
goto found;
|
|
|
|
|
}
|
|
|
|
|
|
2008-10-31 21:10:00 +00:00
|
|
|
/****** HNP AND DAEMONS ONLY ******/
|
|
|
|
|
|
2009-02-06 15:41:53 +00:00
|
|
|
/* if the job family is zero, then this is going to a local slave,
|
|
|
|
|
* so the path is direct
|
|
|
|
|
*/
|
|
|
|
|
if (0 == ORTE_JOB_FAMILY(target->jobid)) {
|
|
|
|
|
ret = target;
|
|
|
|
|
goto found;
|
|
|
|
|
}
|
|
|
|
|
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
/* IF THIS IS FOR A DIFFERENT JOB FAMILY... */
|
|
|
|
|
if (ORTE_JOB_FAMILY(target->jobid) != ORTE_JOB_FAMILY(ORTE_PROC_MY_NAME->jobid)) {
|
|
|
|
|
/* if I am a daemon, route this via the HNP */
|
2009-05-04 11:07:40 +00:00
|
|
|
if (ORTE_PROC_IS_DAEMON) {
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
ret = ORTE_PROC_MY_HNP;
|
|
|
|
|
goto found;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* if I am the HNP or a tool, then I stored a route to
|
|
|
|
|
* this job family, so look it up
|
|
|
|
|
*/
|
2010-07-20 04:47:31 +00:00
|
|
|
jfamily = ORTE_JOB_FAMILY(target->jobid);
|
2010-07-20 08:30:47 +00:00
|
|
|
for (i=0; i < orte_routed_jobfams.size; i++) {
|
|
|
|
|
if (NULL == (jfam = (orte_routed_jobfam_t*)opal_pointer_array_get_item(&orte_routed_jobfams, i))) {
|
2010-07-20 04:47:31 +00:00
|
|
|
continue;
|
|
|
|
|
}
|
|
|
|
|
if (jfam->job_family == jfamily) {
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_routed_base_output,
|
|
|
|
|
"%s routed_binomial: route to %s found",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_JOB_FAMILY_PRINT(target->jobid)));
|
|
|
|
|
ret = &jfam->route;
|
|
|
|
|
goto found;
|
|
|
|
|
}
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
}
|
|
|
|
|
/* not found - so we have no route */
|
|
|
|
|
ret = ORTE_NAME_INVALID;
|
|
|
|
|
goto found;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* THIS CAME FROM OUR OWN JOB FAMILY... */
|
2008-10-31 21:10:00 +00:00
|
|
|
|
2009-01-30 18:31:43 +00:00
|
|
|
/* if we are not using static ports and this is going to the HNP, send direct */
|
|
|
|
|
if (!orte_static_ports &&
|
|
|
|
|
ORTE_PROC_MY_HNP->jobid == target->jobid &&
|
2008-10-31 21:10:00 +00:00
|
|
|
ORTE_PROC_MY_HNP->vpid == target->vpid) {
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_routed_base_output,
|
|
|
|
|
"%s routing not enabled - going direct",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
2008-10-31 21:10:00 +00:00
|
|
|
ret = target;
|
2008-02-28 01:57:57 +00:00
|
|
|
goto found;
|
2007-07-20 01:34:02 +00:00
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
|
2008-10-31 21:10:00 +00:00
|
|
|
daemon.jobid = ORTE_PROC_MY_NAME->jobid;
|
|
|
|
|
/* find out what daemon hosts this proc */
|
|
|
|
|
if (ORTE_VPID_INVALID == (daemon.vpid = orte_ess.proc_get_daemon(target))) {
|
2008-11-18 15:35:50 +00:00
|
|
|
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
|
|
|
|
|
ret = ORTE_NAME_INVALID;
|
2008-02-28 01:57:57 +00:00
|
|
|
goto found;
|
|
|
|
|
}
|
|
|
|
|
|
2008-10-31 21:10:00 +00:00
|
|
|
/* if the daemon is me, then send direct to the target! */
|
|
|
|
|
if (ORTE_PROC_MY_NAME->vpid == daemon.vpid) {
|
|
|
|
|
ret = target;
|
|
|
|
|
} else {
|
|
|
|
|
/* the linear routing tree is trivial - if the vpid is
|
|
|
|
|
* lower than mine, route through my parent, which is
|
|
|
|
|
* at my_vpid-1. If the vpid is higher than mine, then
|
|
|
|
|
* route to my_vpid+1, wrapping around to 0
|
|
|
|
|
*/
|
|
|
|
|
if (daemon.vpid < ORTE_PROC_MY_NAME->vpid) {
|
|
|
|
|
daemon.vpid = ORTE_PROC_MY_NAME->vpid - 1;
|
|
|
|
|
ret = &daemon;
|
|
|
|
|
} else {
|
2009-03-05 21:56:03 +00:00
|
|
|
if (ORTE_PROC_MY_NAME->vpid < orte_process_info.num_procs-1) {
|
2008-10-31 21:10:00 +00:00
|
|
|
daemon.vpid = ORTE_PROC_MY_NAME->vpid + 1;
|
|
|
|
|
} else {
|
|
|
|
|
/* we are at end of chain - wrap around */
|
|
|
|
|
daemon.vpid = 0;
|
|
|
|
|
}
|
|
|
|
|
ret = &daemon;
|
|
|
|
|
}
|
|
|
|
|
}
|
2007-07-20 01:34:02 +00:00
|
|
|
|
|
|
|
|
found:
|
2009-06-24 19:49:06 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
2008-05-05 02:59:09 +00:00
|
|
|
"%s routed_linear_get(%s) --> %s",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_NAME_PRINT(target),
|
|
|
|
|
ORTE_NAME_PRINT(ret)));
|
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
|
|
|
|
2008-03-05 22:44:35 +00:00
|
|
|
return *ret;
|
2007-07-20 01:34:02 +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
|
|
|
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
/* HANDLE ACK MESSAGES FROM AN HNP */
|
|
|
|
|
static void release_ack(int fd, short event, void *data)
|
|
|
|
|
{
|
|
|
|
|
orte_message_event_t *mev = (orte_message_event_t*)data;
|
|
|
|
|
ack_recvd = true;
|
|
|
|
|
OBJ_RELEASE(mev);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void recv_ack(int status, orte_process_name_t* sender,
|
|
|
|
|
opal_buffer_t* buffer, orte_rml_tag_t tag,
|
|
|
|
|
void* cbdata)
|
|
|
|
|
{
|
|
|
|
|
/* don't process this right away - we need to get out of the recv before
|
|
|
|
|
* we process the message as it may ask us to do something that involves
|
|
|
|
|
* more messaging! Instead, setup an event so that the message gets processed
|
|
|
|
|
* as soon as we leave the recv.
|
|
|
|
|
*
|
|
|
|
|
* The macro makes a copy of the buffer, which we release above - the incoming
|
|
|
|
|
* buffer, however, is NOT released here, although its payload IS transferred
|
|
|
|
|
* to the message buffer for later processing
|
|
|
|
|
*/
|
|
|
|
|
ORTE_MESSAGE_EVENT(sender, buffer, tag, release_ack);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2008-03-26 01:00:24 +00:00
|
|
|
static int init_routes(orte_jobid_t job, opal_buffer_t *ndat)
|
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
|
|
|
{
|
2008-05-05 02:59:09 +00:00
|
|
|
/* the linear module routes all proc communications through
|
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
|
|
|
* the local daemon. Daemons must identify which of their
|
|
|
|
|
* daemon-peers is "hosting" the specified recipient and
|
|
|
|
|
* route the message to that daemon. Daemon contact info
|
|
|
|
|
* is handled elsewhere, so all we need to do here is
|
|
|
|
|
* ensure that the procs are told to route through their
|
|
|
|
|
* local daemon, and that daemons are told how to route
|
|
|
|
|
* for each proc
|
|
|
|
|
*/
|
|
|
|
|
int rc;
|
|
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
/* if I am a tool, then I stand alone - there is nothing to do */
|
2009-05-04 11:07:40 +00:00
|
|
|
if (ORTE_PROC_IS_TOOL) {
|
2008-02-28 01:57:57 +00:00
|
|
|
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
|
|
|
/* if I am a daemon or HNP, then I have to extract the routing info for this job
|
|
|
|
|
* from the data sent to me for launch and update the routing tables to
|
|
|
|
|
* point at the daemon for each proc
|
|
|
|
|
*/
|
2009-05-04 11:07:40 +00:00
|
|
|
if (ORTE_PROC_IS_DAEMON) {
|
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
|
|
|
|
2008-06-09 14:53:58 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
2008-05-05 02:59:09 +00:00
|
|
|
"%s routed_linear: init routes for daemon job %s\n\thnp_uri %s",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_JOBID_PRINT(job),
|
2009-03-05 21:56:03 +00:00
|
|
|
(NULL == orte_process_info.my_hnp_uri) ? "NULL" : orte_process_info.my_hnp_uri));
|
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
|
|
|
|
|
|
|
|
if (NULL == ndat) {
|
2008-02-28 01:57:57 +00:00
|
|
|
/* indicates this is being called during orte_init.
|
|
|
|
|
* Get the HNP's name for possible later use
|
|
|
|
|
*/
|
2009-03-05 21:56:03 +00:00
|
|
|
if (NULL == orte_process_info.my_hnp_uri) {
|
2008-02-28 01:57:57 +00:00
|
|
|
/* fatal error */
|
|
|
|
|
ORTE_ERROR_LOG(ORTE_ERR_FATAL);
|
|
|
|
|
return ORTE_ERR_FATAL;
|
|
|
|
|
}
|
|
|
|
|
/* set the contact info into the hash table */
|
2009-03-05 21:56:03 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml.set_contact_info(orte_process_info.my_hnp_uri))) {
|
2008-02-28 01:57:57 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return(rc);
|
|
|
|
|
}
|
Squeeeeeeze the launch message. This is the message sent to the daemons that provides all the data required for launching their local procs. In reorganizing the ODLS framework, I discovered that we were sending a significant amount of unnecessary and repeated data. This commit resolves this by:
1. taking advantage of the fact that we no longer create the launch message via a GPR trigger. In earlier times, we had the GPR create the launch message based on a subscription. In that mode of operation, we could not guarantee the order in which the data was stored in the message - hence, we had no choice but to parse the message in a loop that checked each value against a list of possible "keys" until the corresponding value was found.
Now, however, we construct the message "by hand", so we know precisely what data is in each location in the message. Thus, we no longer need to send the character string "keys" for each data value any more. This represents a rather large savings in the message size - to give you an example, we typically would use a 30-char "key" for a 2-byte data value. As you can see, the overhead can become very large.
2. sending node-specific data only once. Again, because we used to construct the message via subscriptions that were done on a per-proc basis, the data for each node (e.g., the daemon's name, whether or not the node was oversubscribed) would be included in the data for each proc. Thus, the node-specific data was repeated for every proc.
Now that we construct the message "by hand", there is no reason to do this any more. Instead, we can insert the data for a specific node only once, and then provide the per-proc data for that node. We therefore not only save all that extra data in the message, but we also only need to parse the per-node data once.
The savings become significant at scale. Here is a comparison between the revised trunk and the trunk prior to this commit (all data was taken on odin, using openib, 64 nodes, unity message routing, tested with application consisting of mpi_init/mpi_barrier/mpi_finalize, all execution times given in seconds, all launch message sizes in bytes):
Per-node scaling, taken at 1ppn:
#nodes original trunk revised trunk
time size time size
1 0.10 819 0.09 564
2 0.14 1070 0.14 677
3 0.15 1321 0.14 790
4 0.15 1572 0.15 903
8 0.17 2576 0.20 1355
16 0.25 4584 0.21 2259
32 0.28 8600 0.27 4067
64 0.50 16632 0.39 7683
Per-proc scaling, taken at 64 nodes
ppn original trunk revised trunk
time size time size
1 0.50 16669 0.40 7720
2 0.55 32733 0.54 11048
3 0.87 48797 0.81 14376
4 1.0 64861 0.85 17704
Condensing those numbers, it appears we gained:
per-node message size: 251 bytes/node -> 113 bytes/node
per-proc message size: 251 bytes/proc -> 52 bytes/proc
per-job message size: 568 bytes/job -> 399 bytes/job
(job-specific data such as jobid, override oversubscribe flag, total #procs in job, total slots allocated)
The fact that the two pre-commit trunk numbers are the same confirms the fact that each proc was containing the node data as well. It isn't quite the 10x message reduction I had hoped to get, but it is significant and gives much better scaling.
Note that the timing info was, as usual, pretty chaotic - the numbers cited here were typical across several runs taken after the initial one to avoid NFS file positioning influences.
Also note that this commit removes the orte_process_info.vpid_start field and the handful of places that passed that useless value. By definition, all jobs start at vpid=0, so all we were doing is passing "0" around. In fact, many places simply hardwired it to "0" anyway rather than deal with it.
This commit was SVN r16428.
2007-10-11 15:57:26 +00:00
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
/* extract the hnp name and store it */
|
2009-03-05 21:56:03 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml_base_parse_uris(orte_process_info.my_hnp_uri,
|
2008-02-28 01:57:57 +00:00
|
|
|
ORTE_PROC_MY_HNP, NULL))) {
|
Squeeeeeeze the launch message. This is the message sent to the daemons that provides all the data required for launching their local procs. In reorganizing the ODLS framework, I discovered that we were sending a significant amount of unnecessary and repeated data. This commit resolves this by:
1. taking advantage of the fact that we no longer create the launch message via a GPR trigger. In earlier times, we had the GPR create the launch message based on a subscription. In that mode of operation, we could not guarantee the order in which the data was stored in the message - hence, we had no choice but to parse the message in a loop that checked each value against a list of possible "keys" until the corresponding value was found.
Now, however, we construct the message "by hand", so we know precisely what data is in each location in the message. Thus, we no longer need to send the character string "keys" for each data value any more. This represents a rather large savings in the message size - to give you an example, we typically would use a 30-char "key" for a 2-byte data value. As you can see, the overhead can become very large.
2. sending node-specific data only once. Again, because we used to construct the message via subscriptions that were done on a per-proc basis, the data for each node (e.g., the daemon's name, whether or not the node was oversubscribed) would be included in the data for each proc. Thus, the node-specific data was repeated for every proc.
Now that we construct the message "by hand", there is no reason to do this any more. Instead, we can insert the data for a specific node only once, and then provide the per-proc data for that node. We therefore not only save all that extra data in the message, but we also only need to parse the per-node data once.
The savings become significant at scale. Here is a comparison between the revised trunk and the trunk prior to this commit (all data was taken on odin, using openib, 64 nodes, unity message routing, tested with application consisting of mpi_init/mpi_barrier/mpi_finalize, all execution times given in seconds, all launch message sizes in bytes):
Per-node scaling, taken at 1ppn:
#nodes original trunk revised trunk
time size time size
1 0.10 819 0.09 564
2 0.14 1070 0.14 677
3 0.15 1321 0.14 790
4 0.15 1572 0.15 903
8 0.17 2576 0.20 1355
16 0.25 4584 0.21 2259
32 0.28 8600 0.27 4067
64 0.50 16632 0.39 7683
Per-proc scaling, taken at 64 nodes
ppn original trunk revised trunk
time size time size
1 0.50 16669 0.40 7720
2 0.55 32733 0.54 11048
3 0.87 48797 0.81 14376
4 1.0 64861 0.85 17704
Condensing those numbers, it appears we gained:
per-node message size: 251 bytes/node -> 113 bytes/node
per-proc message size: 251 bytes/proc -> 52 bytes/proc
per-job message size: 568 bytes/job -> 399 bytes/job
(job-specific data such as jobid, override oversubscribe flag, total #procs in job, total slots allocated)
The fact that the two pre-commit trunk numbers are the same confirms the fact that each proc was containing the node data as well. It isn't quite the 10x message reduction I had hoped to get, but it is significant and gives much better scaling.
Note that the timing info was, as usual, pretty chaotic - the numbers cited here were typical across several runs taken after the initial one to avoid NFS file positioning influences.
Also note that this commit removes the orte_process_info.vpid_start field and the handful of places that passed that useless value. By definition, all jobs start at vpid=0, so all we were doing is passing "0" around. In fact, many places simply hardwired it to "0" anyway rather than deal with it.
This commit was SVN r16428.
2007-10-11 15:57:26 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
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
|
|
|
|
2008-03-06 15:30:44 +00:00
|
|
|
/* set our lifeline to the the HNP - we will abort if that connection is lost */
|
2008-03-26 01:00:24 +00:00
|
|
|
lifeline = ORTE_PROC_MY_HNP;
|
2008-03-06 15:30:44 +00:00
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
/* daemons will send their contact info back to the HNP as
|
|
|
|
|
* part of the message confirming they are read to go. HNP's
|
|
|
|
|
* load their contact info during orte_init
|
|
|
|
|
*/
|
|
|
|
|
} else {
|
|
|
|
|
/* ndat != NULL means we are getting an update of RML info
|
|
|
|
|
* for the daemons - so update our contact info and routes
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml_base_update_contact_info(ndat))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
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
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
return rc;
|
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
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
|
2008-06-09 14:53:58 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_routed_base_output,
|
2008-05-05 02:59:09 +00:00
|
|
|
"%s routed_linear: completed init routes",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
2008-02-28 01:57:57 +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
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2009-05-04 11:07:40 +00:00
|
|
|
if (ORTE_PROC_IS_HNP) {
|
2008-02-28 01:57:57 +00:00
|
|
|
|
2008-06-09 14:53:58 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
2008-05-05 02:59:09 +00:00
|
|
|
"%s routed_linear: init routes for HNP job %s",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_JOBID_PRINT(job)));
|
2008-02-28 01:57:57 +00:00
|
|
|
|
|
|
|
|
if (NULL == ndat) {
|
2008-03-26 01:00:24 +00:00
|
|
|
/* the HNP has no lifeline */
|
|
|
|
|
lifeline = NULL;
|
2008-02-28 01:57:57 +00:00
|
|
|
} else {
|
2008-02-28 19:58:32 +00:00
|
|
|
/* if this is for my own jobid, then I am getting an update of RML info
|
2008-02-28 01:57:57 +00:00
|
|
|
* for the daemons - so update our contact info and routes
|
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
|
|
|
*/
|
2008-02-28 19:58:32 +00:00
|
|
|
if (ORTE_PROC_MY_NAME->jobid == job) {
|
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml_base_update_contact_info(ndat))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
/* if not, then I need to process the callback */
|
2009-12-17 19:39:53 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_routed_base_process_callback(job, ndat))) {
|
2008-02-28 19:58:32 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
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
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
{ /* MUST BE A PROC */
|
|
|
|
|
/* if ndat != NULL, then this is being invoked by the proc to
|
|
|
|
|
* init a route to a specified process that is outside of our
|
|
|
|
|
* job family. We want that route to go through our HNP, routed via
|
|
|
|
|
* out local daemon - however, we cannot know for
|
|
|
|
|
* certain that the HNP already knows how to talk to the specified
|
|
|
|
|
* procs. For example, in OMPI's publish/subscribe procedures, the
|
|
|
|
|
* DPM framework looks for an mca param containing the global ompi-server's
|
|
|
|
|
* uri. This info will come here so the proc can setup a route to
|
|
|
|
|
* the server - we need to pass the routing info to our HNP
|
|
|
|
|
*/
|
|
|
|
|
if (NULL != ndat) {
|
2010-07-20 08:30:47 +00:00
|
|
|
int rc;
|
2010-07-20 04:22:45 +00:00
|
|
|
opal_buffer_t xfer;
|
|
|
|
|
orte_rml_cmd_flag_t cmd=ORTE_RML_UPDATE_CMD;
|
2008-02-28 01:57:57 +00:00
|
|
|
|
2008-06-09 14:53:58 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
2008-05-05 02:59:09 +00:00
|
|
|
"%s routed_linear: init routes w/non-NULL data",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
2008-02-28 01:57:57 +00:00
|
|
|
|
2009-02-06 15:41:53 +00:00
|
|
|
/* if this is for a job family of zero, then we know that the enclosed
|
|
|
|
|
* procs are local slaves to our daemon. In that case, we can just ignore this
|
|
|
|
|
* as our daemon - given that it had to spawn the local slave - already
|
|
|
|
|
* knows how to talk to them
|
|
|
|
|
*/
|
|
|
|
|
if (0 == ORTE_JOB_FAMILY(job)) {
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
2008-11-18 15:35:50 +00:00
|
|
|
if (ORTE_JOB_FAMILY(ORTE_PROC_MY_NAME->jobid) != ORTE_JOB_FAMILY(job)) {
|
|
|
|
|
/* if this is for a different job family, then we route via our HNP
|
|
|
|
|
* to minimize connection counts to entities such as ompi-server, so
|
|
|
|
|
* start by sending the contact info to the HNP for update
|
|
|
|
|
*/
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
|
|
|
|
"%s routed_linear_init_routes: diff job family - sending update to %s",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_HNP)));
|
2008-11-18 15:35:50 +00:00
|
|
|
|
2010-07-20 04:22:45 +00:00
|
|
|
/* prep the buffer for transmission to the HNP */
|
|
|
|
|
OBJ_CONSTRUCT(&xfer, opal_buffer_t);
|
|
|
|
|
opal_dss.pack(&xfer, &cmd, 1, ORTE_RML_CMD);
|
|
|
|
|
opal_dss.copy_payload(&xfer, ndat);
|
|
|
|
|
|
|
|
|
|
/* save any new connections for use in subsequent connect_accept calls */
|
2010-07-20 08:30:47 +00:00
|
|
|
orte_routed_base_update_hnps(ndat);
|
2010-07-20 04:22:45 +00:00
|
|
|
|
|
|
|
|
if (0 > (rc = orte_rml.send_buffer(ORTE_PROC_MY_HNP, &xfer,
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
ORTE_RML_TAG_RML_INFO_UPDATE, 0))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
2010-07-20 04:22:45 +00:00
|
|
|
OBJ_DESTRUCT(&xfer);
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
return rc;
|
|
|
|
|
}
|
2010-07-20 04:22:45 +00:00
|
|
|
OBJ_DESTRUCT(&xfer);
|
2008-11-18 15:35:50 +00:00
|
|
|
|
|
|
|
|
/* wait right here until the HNP acks the update to ensure that
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
* any subsequent messaging can succeed
|
|
|
|
|
*/
|
|
|
|
|
ack_recvd = false;
|
|
|
|
|
rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_UPDATE_ROUTE_ACK,
|
|
|
|
|
ORTE_RML_NON_PERSISTENT, recv_ack, NULL);
|
|
|
|
|
|
|
|
|
|
ORTE_PROGRESSED_WAIT(ack_recvd, 0, 1);
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
|
|
|
|
"%s routed_linear_init_routes: ack recvd",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
|
2008-11-18 15:35:50 +00:00
|
|
|
/* our get_route function automatically routes all messages for
|
|
|
|
|
* other job families via the HNP, so nothing more to do here
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
*/
|
|
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
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
|
|
|
}
|
|
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
/* if ndat=NULL, then we are being called during orte_init. In this
|
|
|
|
|
* case, we need to setup a few critical pieces of info
|
|
|
|
|
*/
|
|
|
|
|
|
2008-06-09 14:53:58 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
2008-05-05 02:59:09 +00:00
|
|
|
"%s routed_linear: init routes for proc job %s\n\thnp_uri %s\n\tdaemon uri %s",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), ORTE_JOBID_PRINT(job),
|
2009-03-05 21:56:03 +00:00
|
|
|
(NULL == orte_process_info.my_hnp_uri) ? "NULL" : orte_process_info.my_hnp_uri,
|
|
|
|
|
(NULL == orte_process_info.my_daemon_uri) ? "NULL" : orte_process_info.my_daemon_uri));
|
2008-02-28 01:57:57 +00:00
|
|
|
|
2009-03-05 21:56:03 +00:00
|
|
|
if (NULL == orte_process_info.my_daemon_uri) {
|
2008-02-28 01:57:57 +00:00
|
|
|
/* in this module, we absolutely MUST have this information - if
|
|
|
|
|
* we didn't get it, then error out
|
|
|
|
|
*/
|
2008-06-09 14:53:58 +00:00
|
|
|
opal_output(0, "%s ERROR: Failed to identify the local daemon's URI",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
|
2008-06-09 14:53:58 +00:00
|
|
|
opal_output(0, "%s ERROR: This is a fatal condition when the linear router",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
|
2008-06-09 14:53:58 +00:00
|
|
|
opal_output(0, "%s ERROR: has been selected - either select the unity router",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
|
2008-06-09 14:53:58 +00:00
|
|
|
opal_output(0, "%s ERROR: or ensure that the local daemon info is provided",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
|
2008-02-28 01:57:57 +00:00
|
|
|
return ORTE_ERR_FATAL;
|
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
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
|
|
|
|
|
/* we have to set the HNP's name, even though we won't route messages directly
|
|
|
|
|
* to it. This is required to ensure that we -do- send messages to the correct
|
|
|
|
|
* HNP name
|
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
|
|
|
*/
|
2009-03-05 21:56:03 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml_base_parse_uris(orte_process_info.my_hnp_uri,
|
2008-02-28 01:57:57 +00:00
|
|
|
ORTE_PROC_MY_HNP, NULL))) {
|
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_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
/* Set the contact info in the RML - this won't actually establish
|
|
|
|
|
* the connection, but just tells the RML how to reach the daemon
|
|
|
|
|
* if/when we attempt to send to it
|
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
|
|
|
*/
|
2009-03-05 21:56:03 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml.set_contact_info(orte_process_info.my_daemon_uri))) {
|
2008-02-28 01:57:57 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return(rc);
|
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
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
/* extract the daemon's name so we can update the routing table */
|
2009-03-05 21:56:03 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml_base_parse_uris(orte_process_info.my_daemon_uri,
|
2008-02-28 01:57:57 +00:00
|
|
|
ORTE_PROC_MY_DAEMON, NULL))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
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
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
|
2008-03-06 15:30:44 +00:00
|
|
|
/* set our lifeline to the local daemon - we will abort if this connection is lost */
|
2008-03-26 01:00:24 +00:00
|
|
|
lifeline = ORTE_PROC_MY_DAEMON;
|
2008-03-06 15:30:44 +00:00
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
/* register ourselves -this sends a message to the daemon (warming up that connection)
|
|
|
|
|
* and sends our contact info to the HNP when all local procs have reported
|
|
|
|
|
*
|
|
|
|
|
* NOTE: it may seem odd that we send our contact info to the HNP - after all,
|
|
|
|
|
* the HNP doesn't really need to know how to talk to us directly if we are
|
|
|
|
|
* using this routing method. However, this is good for two reasons:
|
|
|
|
|
*
|
|
|
|
|
* (1) some debuggers and/or tools may need RML contact
|
|
|
|
|
* info to set themselves up
|
|
|
|
|
*
|
|
|
|
|
* (2) doing so allows the HNP to "block" in a dynamic launch
|
|
|
|
|
* until all procs are reported running, thus ensuring that no communication
|
|
|
|
|
* is attempted until the overall ORTE system knows how to talk to everyone -
|
|
|
|
|
* otherwise, the system can just hang.
|
|
|
|
|
*/
|
2008-04-30 19:49:53 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_routed_base_register_sync(true))) {
|
2008-02-28 01:57:57 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
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
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
/* no answer is expected or coming */
|
|
|
|
|
|
|
|
|
|
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
|
|
|
}
|
|
|
|
|
}
|
2008-03-26 01:00:24 +00:00
|
|
|
|
|
|
|
|
static int route_lost(const orte_process_name_t *route)
|
|
|
|
|
{
|
|
|
|
|
/* if we lose the connection to the lifeline and we are NOT already,
|
|
|
|
|
* in finalize, tell the OOB to abort.
|
|
|
|
|
* NOTE: we cannot call abort from here as the OOB needs to first
|
|
|
|
|
* release a thread-lock - otherwise, we will hang!!
|
|
|
|
|
*/
|
|
|
|
|
if (!orte_finalizing &&
|
|
|
|
|
NULL != lifeline &&
|
|
|
|
|
OPAL_EQUAL == orte_util_compare_name_fields(ORTE_NS_CMP_ALL, route, lifeline)) {
|
2008-06-09 14:53:58 +00:00
|
|
|
opal_output(0, "%s routed:linear: Connection to lifeline %s lost",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_NAME_PRINT(lifeline));
|
2008-03-26 01:00:24 +00:00
|
|
|
return ORTE_ERR_FATAL;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* we don't care about this one, so return success */
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2008-05-05 02:59:09 +00:00
|
|
|
static bool route_is_defined(const orte_process_name_t *target)
|
|
|
|
|
{
|
2008-10-31 21:10:00 +00:00
|
|
|
/* find out what daemon hosts this proc */
|
|
|
|
|
if (ORTE_VPID_INVALID == orte_ess.proc_get_daemon((orte_process_name_t*)target)) {
|
|
|
|
|
return false;
|
2008-05-05 19:12:41 +00:00
|
|
|
}
|
2008-10-31 21:10:00 +00:00
|
|
|
|
|
|
|
|
return true;
|
2008-05-05 02:59:09 +00:00
|
|
|
}
|
|
|
|
|
|
2008-05-05 22:32:25 +00:00
|
|
|
|
2008-12-20 23:23:11 +00:00
|
|
|
static int set_lifeline(orte_process_name_t *proc)
|
|
|
|
|
{
|
|
|
|
|
/* we have to copy the proc data because there is no
|
|
|
|
|
* guarantee that it will be preserved
|
|
|
|
|
*/
|
|
|
|
|
local_lifeline.jobid = proc->jobid;
|
|
|
|
|
local_lifeline.vpid = proc->vpid;
|
|
|
|
|
lifeline = &local_lifeline;
|
|
|
|
|
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
2008-05-05 02:59:09 +00:00
|
|
|
static int update_routing_tree(void)
|
|
|
|
|
{
|
2008-06-20 03:26:13 +00:00
|
|
|
/* if I am anything other than a daemon or the HNP, this
|
|
|
|
|
* is a meaningless command as I am not allowed to route
|
|
|
|
|
*/
|
2009-05-04 11:07:40 +00:00
|
|
|
if (!ORTE_PROC_IS_DAEMON && !ORTE_PROC_IS_HNP) {
|
2008-06-20 03:26:13 +00:00
|
|
|
return ORTE_ERR_NOT_SUPPORTED;
|
|
|
|
|
}
|
|
|
|
|
|
2008-05-05 22:32:25 +00:00
|
|
|
/* nothing to do here as the routing tree is fixed */
|
2008-05-05 02:59:09 +00:00
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
2009-01-27 19:13:56 +00:00
|
|
|
static orte_vpid_t get_routing_tree(opal_list_t *children)
|
2008-05-05 02:59:09 +00:00
|
|
|
{
|
2009-01-27 19:13:56 +00:00
|
|
|
orte_routed_tree_t *nm;
|
|
|
|
|
orte_vpid_t v;
|
2008-05-05 22:32:25 +00:00
|
|
|
|
2008-06-20 03:26:13 +00:00
|
|
|
/* if I am anything other than a daemon or the HNP, this
|
|
|
|
|
* is a meaningless command as I am not allowed to route
|
|
|
|
|
*/
|
2009-05-04 11:07:40 +00:00
|
|
|
if (!ORTE_PROC_IS_DAEMON && !ORTE_PROC_IS_HNP) {
|
2008-06-20 03:26:13 +00:00
|
|
|
return ORTE_VPID_INVALID;
|
|
|
|
|
}
|
|
|
|
|
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
/* the linear routing tree consists of a chain of daemons
|
2009-03-05 21:56:03 +00:00
|
|
|
* extending from the HNP to orte_process_info.num_procs-1.
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
* Accordingly, my child is just the my_vpid+1 daemon
|
2008-05-05 22:32:25 +00:00
|
|
|
*/
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
if (NULL != children &&
|
2009-03-05 21:56:03 +00:00
|
|
|
ORTE_PROC_MY_NAME->vpid < orte_process_info.num_procs-1) {
|
2009-01-27 19:13:56 +00:00
|
|
|
/* my child is just the vpid+1 daemon */
|
|
|
|
|
nm = OBJ_NEW(orte_routed_tree_t);
|
2009-03-05 21:56:03 +00:00
|
|
|
opal_bitmap_init(&nm->relatives, orte_process_info.num_procs);
|
2009-01-27 19:13:56 +00:00
|
|
|
nm->vpid = ORTE_PROC_MY_NAME->vpid + 1;
|
|
|
|
|
/* my relatives are everyone above that point */
|
2009-03-05 21:56:03 +00:00
|
|
|
for (v=nm->vpid+1; v < orte_process_info.num_procs; v++) {
|
2009-01-27 19:13:56 +00:00
|
|
|
opal_bitmap_set_bit(&nm->relatives, v);
|
|
|
|
|
}
|
|
|
|
|
opal_list_append(children, &nm->super);
|
2008-05-06 20:16:17 +00:00
|
|
|
}
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
|
2009-05-04 11:07:40 +00:00
|
|
|
if (ORTE_PROC_IS_HNP) {
|
Repair the MPI-2 dynamic operations. This includes:
1. repair of the linear and direct routed modules
2. repair of the ompi/pubsub/orte module to correctly init routes to the ompi-server, and correctly handle failure to correctly parse the provided ompi-server URI
3. modification of orterun to accept both "file" and "FILE" for designating where the ompi-server URI is to be found - purely a convenience feature
4. resolution of a message ordering problem during the connect/accept handshake that allowed the "send-first" proc to attempt to send to the "recv-first" proc before the HNP had actually updated its routes.
Let this be a further reminder to all - message ordering is NOT guaranteed in the OOB
5. Repair the ompi/dpm/orte module to correctly init routes during connect/accept.
Reminder to all: messages sent to procs in another job family (i.e., started by a different mpirun) are ALWAYS routed through the respective HNPs. As per the comments in orte/routed, this is REQUIRED to maintain connect/accept (where only the root proc on each side is capable of init'ing the routes), allow communication between mpirun's using different routing modules, and to minimize connections on tools such as ompi-server. It is all taken care of "under the covers" by the OOB to ensure that a route back to the sender is maintained, even when the different mpirun's are using different routed modules.
6. corrections in the orte/odls to ensure proper identification of daemons participating in a dynamic launch
7. corrections in build/nidmap to support update of an existing nidmap during dynamic launch
8. corrected implementation of the update_arch function in the ESS, along with consolidation of a number of ESS operations into base functions for easier maintenance. The ability to support info from multiple jobs was added, although we don't currently do so - this will come later to support further fault recovery strategies
9. minor updates to several functions to remove unnecessary and/or no longer used variables and envar's, add some debugging output, etc.
10. addition of a new macro ORTE_PROC_IS_DAEMON that resolves to true if the provided proc is a daemon
There is still more cleanup to be done for efficiency, but this at least works.
Tested on single-node Mac, multi-node SLURM via odin. Tests included connect/accept, publish/lookup/unpublish, comm_spawn, comm_spawn_multiple, and singleton comm_spawn.
Fixes ticket #1256
This commit was SVN r18804.
2008-07-03 17:53:37 +00:00
|
|
|
/* the parent of the HNP is invalid */
|
|
|
|
|
return ORTE_VPID_INVALID;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* my parent is the my_vpid-1 daemon */
|
|
|
|
|
return (ORTE_PROC_MY_NAME->vpid - 1);
|
2008-05-05 02:59:09 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2008-10-02 20:08:27 +00:00
|
|
|
static int get_wireup_info(opal_buffer_t *buf)
|
2008-03-26 01:00:24 +00:00
|
|
|
{
|
|
|
|
|
int rc;
|
2010-07-20 08:30:47 +00:00
|
|
|
int i;
|
|
|
|
|
orte_routed_jobfam_t *jfam;
|
|
|
|
|
|
2010-07-20 04:22:45 +00:00
|
|
|
if (ORTE_PROC_IS_HNP) {
|
|
|
|
|
/* if we are not using static ports, then we need to share the
|
|
|
|
|
* comm info - otherwise, just return
|
|
|
|
|
*/
|
|
|
|
|
if (orte_static_ports) {
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
2008-06-20 03:26:13 +00:00
|
|
|
|
2010-07-20 04:22:45 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml_base_get_contact_info(ORTE_PROC_MY_NAME->jobid, buf))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
|
|
|
|
return rc;
|
2008-04-30 19:49:53 +00:00
|
|
|
}
|
|
|
|
|
|
2010-07-20 04:22:45 +00:00
|
|
|
/* if I am an application, this is occurring during connect_accept.
|
|
|
|
|
* We need to return the stored information of other HNPs we
|
|
|
|
|
* know about, if any
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_PROC_IS_APP) {
|
2010-07-20 08:30:47 +00:00
|
|
|
for (i=0; i < orte_routed_jobfams.size; i++) {
|
|
|
|
|
if (NULL != (jfam = (orte_routed_jobfam_t*)opal_pointer_array_get_item(&orte_routed_jobfams, i))) {
|
|
|
|
|
opal_dss.pack(buf, &(jfam->hnp_uri), 1, OPAL_STRING);
|
|
|
|
|
}
|
2010-07-20 04:22:45 +00:00
|
|
|
}
|
2010-07-20 08:30:47 +00:00
|
|
|
return ORTE_SUCCESS;
|
2008-03-26 01:00:24 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
2008-04-23 00:17:12 +00:00
|
|
|
|
2010-07-17 21:03:27 +00:00
|
|
|
static size_t num_routes(void)
|
|
|
|
|
{
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2010-03-12 23:57:50 +00:00
|
|
|
#if OPAL_ENABLE_FT_CR == 1
|
2008-05-05 02:59:09 +00:00
|
|
|
static int linear_ft_event(int state)
|
2008-04-23 00:17:12 +00:00
|
|
|
{
|
|
|
|
|
int ret, exit_status = ORTE_SUCCESS;
|
|
|
|
|
|
|
|
|
|
/******** Checkpoint Prep ********/
|
|
|
|
|
if(OPAL_CRS_CHECKPOINT == state) {
|
|
|
|
|
}
|
|
|
|
|
/******** Continue Recovery ********/
|
|
|
|
|
else if (OPAL_CRS_CONTINUE == state ) {
|
|
|
|
|
}
|
|
|
|
|
/******** Restart Recovery ********/
|
|
|
|
|
else if (OPAL_CRS_RESTART == state ) {
|
|
|
|
|
/*
|
|
|
|
|
* Re-exchange the routes
|
|
|
|
|
*/
|
|
|
|
|
if (ORTE_SUCCESS != (ret = orte_routed.init_routes(ORTE_PROC_MY_NAME->jobid, NULL))) {
|
|
|
|
|
exit_status = ret;
|
|
|
|
|
goto cleanup;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
else if (OPAL_CRS_TERM == state ) {
|
|
|
|
|
/* Nothing */
|
|
|
|
|
}
|
|
|
|
|
else {
|
|
|
|
|
/* Error state = Nothing */
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
cleanup:
|
|
|
|
|
return exit_status;
|
|
|
|
|
}
|
|
|
|
|
#endif
|