2004-11-22 00:37:56 +00:00
|
|
|
/*
|
2005-11-05 19:57:48 +00:00
|
|
|
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
|
|
|
|
|
* University Research and Technology
|
|
|
|
|
* Corporation. All rights reserved.
|
|
|
|
|
* Copyright (c) 2004-2005 The University of Tennessee and The University
|
|
|
|
|
* of Tennessee Research Foundation. All rights
|
|
|
|
|
* reserved.
|
2006-02-07 03:32:36 +00:00
|
|
|
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
|
2004-11-28 20:09:25 +00:00
|
|
|
* University of Stuttgart. All rights reserved.
|
2005-03-24 12:43:37 +00:00
|
|
|
* Copyright (c) 2004-2005 The Regents of the University of California.
|
|
|
|
|
* All rights reserved.
|
2007-05-31 18:27:54 +00:00
|
|
|
* Copyright (c) 2007 Sun Microsystems, Inc. All rights reserved.
|
2004-11-22 01:38:40 +00:00
|
|
|
* $COPYRIGHT$
|
2006-02-07 03:32:36 +00:00
|
|
|
*
|
2004-11-22 01:38:40 +00:00
|
|
|
* Additional copyrights may follow
|
2006-02-07 03:32:36 +00:00
|
|
|
*
|
2004-11-22 00:37:56 +00:00
|
|
|
* $HEADER$
|
|
|
|
|
*/
|
|
|
|
|
|
2006-02-12 01:33:29 +00:00
|
|
|
#include "orte_config.h"
|
2008-02-28 01:57:57 +00:00
|
|
|
#include "orte/constants.h"
|
|
|
|
|
#include "orte/types.h"
|
2006-11-03 18:55:05 +00:00
|
|
|
|
2004-11-20 19:12:43 +00:00
|
|
|
#include <string.h>
|
2008-12-09 23:49:02 +00:00
|
|
|
#include <fcntl.h>
|
2004-11-20 19:12:43 +00:00
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
#include "opal/dss/dss.h"
|
2008-12-09 23:49:02 +00:00
|
|
|
#include "opal/runtime/opal.h"
|
2010-05-17 23:08:56 +00:00
|
|
|
#include "opal/util/opal_sos.h"
|
2008-06-18 22:17:53 +00:00
|
|
|
|
2006-02-12 01:33:29 +00:00
|
|
|
#include "orte/mca/errmgr/errmgr.h"
|
2008-06-18 22:17:53 +00:00
|
|
|
#include "orte/mca/ess/ess.h"
|
2009-01-27 19:13:56 +00:00
|
|
|
#include "orte/mca/odls/base/base.h"
|
2009-04-29 00:49:23 +00:00
|
|
|
#include "orte/mca/odls/odls_types.h"
|
Bring in the code for routing xcast stage gate messages via the local orteds. This code is inactive unless you specifically request it via an mca param oob_xcast_mode (can be set to "linear" or "direct"). Direct mode is the old standard method where we send messages directly to each MPI process. Linear mode sends the xcast message via the orteds, with the HNP sending the message to each orted directly.
There is a binomial algorithm in the code (i.e., the HNP would send to a subset of the orteds, which then relay it on according to the typical log-2 algo), but that has a bug in it so the code won't let you select it even if you tried (and the mca param doesn't show, so you'd *really* have to try).
This also involved a slight change to the oob.xcast API, so propagated that as required.
Note: this has *only* been tested on rsh, SLURM, and Bproc environments (now that it has been transferred to the OMPI trunk, I'll need to re-test it [only done rsh so far]). It should work fine on any environment that uses the ORTE daemons - anywhere else, you are on your own... :-)
Also, correct a mistake where the orte_debug_flag was declared an int, but the mca param was set as a bool. Move the storage for that flag to the orte/runtime/params.c and orte/runtime/params.h files appropriately.
This commit was SVN r14475.
2007-04-23 18:41:04 +00:00
|
|
|
#include "orte/mca/rml/rml.h"
|
2009-02-14 02:26:12 +00:00
|
|
|
#include "orte/mca/rml/rml_types.h"
|
2009-01-27 19:13:56 +00:00
|
|
|
#include "orte/mca/routed/routed.h"
|
2008-02-28 01:57:57 +00:00
|
|
|
#include "orte/util/name_fns.h"
|
2008-06-18 22:17:53 +00:00
|
|
|
#include "orte/util/show_help.h"
|
|
|
|
|
#include "orte/util/proc_info.h"
|
2009-01-07 15:00:26 +00:00
|
|
|
#include "orte/util/nidmap.h"
|
2008-02-28 01:57:57 +00:00
|
|
|
#include "orte/orted/orted.h"
|
2008-02-28 19:58:32 +00:00
|
|
|
#include "orte/runtime/orte_wait.h"
|
2008-06-18 22:17:53 +00:00
|
|
|
#include "orte/runtime/orte_globals.h"
|
2004-11-20 19:12:43 +00:00
|
|
|
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
#include "orte/mca/grpcomm/base/base.h"
|
2007-07-20 01:34:02 +00:00
|
|
|
#include "grpcomm_basic.h"
|
2004-11-20 19:12:43 +00:00
|
|
|
|
2008-02-28 01:57:57 +00:00
|
|
|
|
2008-03-24 20:50:31 +00:00
|
|
|
/* Static API's */
|
|
|
|
|
static int init(void);
|
|
|
|
|
static void finalize(void);
|
|
|
|
|
static int xcast(orte_jobid_t job,
|
|
|
|
|
opal_buffer_t *buffer,
|
|
|
|
|
orte_rml_tag_t tag);
|
2010-02-25 01:11:29 +00:00
|
|
|
static int basic_allgather(opal_buffer_t *sbuf, opal_buffer_t *rbuf);
|
|
|
|
|
static int basic_barrier(void);
|
|
|
|
|
static int basic_onesided_barrier(void);
|
2008-06-18 22:17:53 +00:00
|
|
|
static int modex(opal_list_t *procs);
|
2008-12-09 23:49:02 +00:00
|
|
|
static int set_proc_attr(const char *attr_name, const void *data, size_t size);
|
2009-01-07 15:00:26 +00:00
|
|
|
static int get_proc_attr(const orte_process_name_t proc,
|
|
|
|
|
const char * attribute_name, void **val,
|
|
|
|
|
size_t *size);
|
2008-03-24 20:50:31 +00:00
|
|
|
|
|
|
|
|
/* Module def */
|
|
|
|
|
orte_grpcomm_base_module_t orte_grpcomm_basic_module = {
|
|
|
|
|
init,
|
|
|
|
|
finalize,
|
|
|
|
|
xcast,
|
2010-02-25 01:11:29 +00:00
|
|
|
basic_allgather,
|
2008-03-24 20:50:31 +00:00
|
|
|
orte_grpcomm_base_allgather_list,
|
2010-02-25 01:11:29 +00:00
|
|
|
basic_barrier,
|
|
|
|
|
basic_onesided_barrier,
|
2008-12-09 23:49:02 +00:00
|
|
|
set_proc_attr,
|
2009-01-07 15:00:26 +00:00
|
|
|
get_proc_attr,
|
2008-06-18 22:17:53 +00:00
|
|
|
modex,
|
2008-03-24 20:50:31 +00:00
|
|
|
orte_grpcomm_base_purge_proc_attrs
|
|
|
|
|
};
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
/* Local variables */
|
|
|
|
|
static orte_grpcomm_collective_t barrier, allgather, onesided_barrier;
|
|
|
|
|
|
2009-01-07 15:00:26 +00:00
|
|
|
static bool recv_on;
|
|
|
|
|
static opal_buffer_t *profile_buf=NULL;
|
2009-01-27 19:13:56 +00:00
|
|
|
static int profile_fd = -1;
|
|
|
|
|
static void profile_recv(int status, orte_process_name_t* sender,
|
|
|
|
|
opal_buffer_t* buffer, orte_rml_tag_t tag,
|
|
|
|
|
void* cbdata);
|
2008-12-09 23:49:02 +00:00
|
|
|
|
2008-03-24 20:50:31 +00:00
|
|
|
/**
|
|
|
|
|
* Initialize the module
|
|
|
|
|
*/
|
|
|
|
|
static int init(void)
|
|
|
|
|
{
|
|
|
|
|
int rc;
|
2008-12-09 23:49:02 +00:00
|
|
|
int value;
|
2008-03-24 20:50:31 +00:00
|
|
|
|
2009-01-07 15:00:26 +00:00
|
|
|
mca_base_param_reg_int_name("orte", "grpcomm_recv_on",
|
|
|
|
|
"Turn on grpcomm recv for profile purposes",
|
|
|
|
|
true, false,
|
|
|
|
|
(int) false, &value);
|
|
|
|
|
recv_on = OPAL_INT_TO_BOOL(value);
|
|
|
|
|
|
2008-03-24 20:50:31 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_grpcomm_base_modex_init())) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
2008-12-09 23:49:02 +00:00
|
|
|
|
2009-05-04 11:07:40 +00:00
|
|
|
if (opal_profile && ORTE_PROC_IS_MPI) {
|
2009-01-07 15:00:26 +00:00
|
|
|
/* if I am an MPI application proc, then create a buffer
|
|
|
|
|
* to pack all my attributes in */
|
|
|
|
|
profile_buf = OBJ_NEW(opal_buffer_t);
|
|
|
|
|
/* seed it with the node name */
|
2009-03-05 21:56:03 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(profile_buf, &orte_process_info.nodename, 1, OPAL_STRING))) {
|
2009-01-07 15:00:26 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
|
|
|
|
}
|
2008-12-09 23:49:02 +00:00
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
/* setup global variables */
|
|
|
|
|
OBJ_CONSTRUCT(&barrier, orte_grpcomm_collective_t);
|
|
|
|
|
OBJ_CONSTRUCT(&allgather, orte_grpcomm_collective_t);
|
|
|
|
|
OBJ_CONSTRUCT(&onesided_barrier, orte_grpcomm_collective_t);
|
|
|
|
|
|
2009-05-04 11:07:40 +00:00
|
|
|
if (ORTE_PROC_IS_HNP && recv_on) {
|
2009-01-27 19:13:56 +00:00
|
|
|
/* open the profile file for writing */
|
|
|
|
|
if (NULL == opal_profile_file) {
|
|
|
|
|
/* no file specified - we will just ignore any incoming data */
|
|
|
|
|
profile_fd = -1;
|
|
|
|
|
} else {
|
|
|
|
|
profile_fd = open(opal_profile_file, O_CREAT|O_RDWR|O_TRUNC, 0644);
|
|
|
|
|
if (profile_fd < 0) {
|
|
|
|
|
/* couldn't be opened */
|
|
|
|
|
ORTE_ERROR_LOG(ORTE_ERR_FILE_OPEN_FAILURE);
|
|
|
|
|
return ORTE_ERR_FILE_OPEN_FAILURE;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD,
|
|
|
|
|
ORTE_RML_TAG_GRPCOMM_PROFILE,
|
|
|
|
|
ORTE_RML_NON_PERSISTENT,
|
|
|
|
|
profile_recv,
|
|
|
|
|
NULL))) {
|
2008-12-09 23:49:02 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
2009-01-27 19:13:56 +00:00
|
|
|
}
|
2008-12-09 23:49:02 +00:00
|
|
|
}
|
|
|
|
|
|
2009-01-27 19:13:56 +00:00
|
|
|
/* if we are a daemon or the hnp, we need to post a
|
|
|
|
|
* recv to catch any collective operations
|
|
|
|
|
*/
|
2009-05-04 11:07:40 +00:00
|
|
|
if (ORTE_PROC_IS_DAEMON || ORTE_PROC_IS_HNP) {
|
2009-01-27 19:13:56 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD,
|
|
|
|
|
ORTE_RML_TAG_DAEMON_COLLECTIVE,
|
|
|
|
|
ORTE_RML_NON_PERSISTENT,
|
2010-02-25 01:11:29 +00:00
|
|
|
orte_grpcomm_base_daemon_coll_recv,
|
2009-01-27 19:13:56 +00:00
|
|
|
NULL))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return rc;
|
2008-03-24 20:50:31 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
|
* Finalize the module
|
|
|
|
|
*/
|
|
|
|
|
static void finalize(void)
|
|
|
|
|
{
|
2009-01-07 15:00:26 +00:00
|
|
|
opal_byte_object_t bo, *boptr;
|
|
|
|
|
opal_buffer_t profile;
|
|
|
|
|
|
2008-03-24 20:50:31 +00:00
|
|
|
orte_grpcomm_base_modex_finalize();
|
2008-12-09 23:49:02 +00:00
|
|
|
|
2009-05-04 11:07:40 +00:00
|
|
|
if (opal_profile && ORTE_PROC_IS_MPI) {
|
2009-01-07 15:00:26 +00:00
|
|
|
/* if I am an MPI proc, send my buffer to the collector */
|
|
|
|
|
boptr = &bo;
|
|
|
|
|
opal_dss.unload(profile_buf, (void**)&boptr->bytes, &boptr->size);
|
|
|
|
|
OBJ_RELEASE(profile_buf);
|
|
|
|
|
/* store it as a single object */
|
|
|
|
|
OBJ_CONSTRUCT(&profile, opal_buffer_t);
|
|
|
|
|
opal_dss.pack(&profile, &boptr, 1, OPAL_BYTE_OBJECT);
|
|
|
|
|
/* send the buffer */
|
|
|
|
|
orte_rml.send_buffer(ORTE_PROC_MY_HNP, &profile, ORTE_RML_TAG_GRPCOMM_PROFILE, 0);
|
|
|
|
|
/* done with buffer */
|
|
|
|
|
OBJ_DESTRUCT(&profile);
|
2008-12-09 23:49:02 +00:00
|
|
|
}
|
2009-01-07 15:00:26 +00:00
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
/* destruct the globals */
|
|
|
|
|
OBJ_DESTRUCT(&barrier);
|
|
|
|
|
OBJ_DESTRUCT(&allgather);
|
|
|
|
|
OBJ_DESTRUCT(&onesided_barrier);
|
|
|
|
|
|
2009-05-04 11:07:40 +00:00
|
|
|
if (ORTE_PROC_IS_HNP && recv_on) {
|
2009-01-07 15:00:26 +00:00
|
|
|
/* if we are profiling and I am the HNP, then stop the
|
|
|
|
|
* profiling receive
|
|
|
|
|
*/
|
2009-01-27 19:13:56 +00:00
|
|
|
orte_rml.recv_cancel(ORTE_NAME_WILDCARD, ORTE_RML_TAG_GRPCOMM_PROFILE);
|
|
|
|
|
if (0 <= profile_fd) {
|
|
|
|
|
close(profile_fd);
|
|
|
|
|
profile_fd = -1;
|
|
|
|
|
}
|
2009-01-07 15:00:26 +00:00
|
|
|
}
|
2009-01-27 19:13:56 +00:00
|
|
|
|
|
|
|
|
/* if we are a daemon or the hnp, we need to cancel the
|
|
|
|
|
* recv we posted
|
|
|
|
|
*/
|
2009-05-04 11:07:40 +00:00
|
|
|
if (ORTE_PROC_IS_DAEMON || ORTE_PROC_IS_HNP) {
|
2009-01-27 19:13:56 +00:00
|
|
|
orte_rml.recv_cancel(ORTE_NAME_WILDCARD, ORTE_RML_TAG_DAEMON_COLLECTIVE);
|
|
|
|
|
}
|
2008-03-24 20:50:31 +00:00
|
|
|
}
|
|
|
|
|
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
/**
|
|
|
|
|
* A "broadcast-like" function to a job's processes.
|
|
|
|
|
* @param jobid The job whose processes are to receive the message
|
|
|
|
|
* @param buffer The data to broadcast
|
|
|
|
|
*/
|
2006-12-01 22:30:39 +00:00
|
|
|
|
2007-07-20 01:34:02 +00:00
|
|
|
static int xcast(orte_jobid_t job,
|
2008-02-28 01:57:57 +00:00
|
|
|
opal_buffer_t *buffer,
|
2007-07-20 01:34:02 +00:00
|
|
|
orte_rml_tag_t tag)
|
2006-12-01 22:30:39 +00:00
|
|
|
{
|
2006-12-03 00:19:11 +00:00
|
|
|
int rc = ORTE_SUCCESS;
|
2008-05-05 22:32:25 +00:00
|
|
|
opal_buffer_t buf;
|
2006-12-01 22:30:39 +00:00
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
|
2008-03-17 19:34:36 +00:00
|
|
|
"%s grpcomm:xcast sent to job %s tag %ld",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_JOBID_PRINT(job), (long)tag));
|
2007-07-20 01:34:02 +00:00
|
|
|
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
/* if there is no message to send, then just return ok */
|
|
|
|
|
if (NULL == buffer) {
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
These changes were mostly captured in a prior RFC (except for #2 below) and are aimed specifically at improving startup performance and setting up the remaining modifications described in that RFC.
The commit has been tested for C/R and Cray operations, and on Odin (SLURM, rsh) and RoadRunner (TM). I tried to update all environments, but obviously could not test them. I know that Windows needs some work, and have highlighted what is know to be needed in the odls process component.
This represents a lot of work by Brian, Tim P, Josh, and myself, with much advice from Jeff and others. For posterity, I have appended a copy of the email describing the work that was done:
As we have repeatedly noted, the modex operation in MPI_Init is the single greatest consumer of time during startup. To-date, we have executed that operation as an ORTE stage gate that held the process until a startup message containing all required modex (and OOB contact info - see #3 below) info could be sent to it. Each process would send its data to the HNP's registry, which assembled and sent the message when all processes had reported in.
In addition, ORTE had taken responsibility for monitoring process status as it progressed through a series of "stage gates". The process reported its status at each gate, and ORTE would then send a "release" message once all procs had reported in.
The incoming changes revamp these procedures in three ways:
1. eliminating the ORTE stage gate system and cleanly delineating responsibility between the OMPI and ORTE layers for MPI init/finalize. The modex stage gate (STG1) has been replaced by a collective operation in the modex itself that performs an allgather on the required modex info. The allgather is implemented using the orte_grpcomm framework since the BTL's are not active at that point. At the moment, the grpcomm framework only has a "basic" component analogous to OMPI's "basic" coll framework - I would recommend that the MPI team create additional, more advanced components to improve performance of this step.
The other stage gates have been replaced by orte_grpcomm barrier functions. We tried to use MPI barriers instead (since the BTL's are active at that point), but - as we discussed on the telecon - these are not currently true barriers so the job would hang when we fell through while messages were still in process. Note that the grpcomm barrier doesn't actually resolve that problem, but Brian has pointed out that we are unlikely to ever see it violated. Again, you might want to spend a little time on an advanced barrier algorithm as the one in "basic" is very simplistic.
Summarizing this change: ORTE no longer tracks process state nor has direct responsibility for synchronizing jobs. This is now done via collective operations within the MPI layer, albeit using ORTE collective communication services. I -strongly- urge the MPI team to implement advanced collective algorithms to improve the performance of this critical procedure.
2. reducing the volume of data exchanged during modex. Data in the modex consisted of the process name, the name of the node where that process is located (expressed as a string), plus a string representation of all contact info. The nodename was required in order for the modex to determine if the process was local or not - in addition, some people like to have it to print pretty error messages when a connection failed.
The size of this data has been reduced in three ways:
(a) reducing the size of the process name itself. The process name consisted of two 32-bit fields for the jobid and vpid. This is far larger than any current system, or system likely to exist in the near future, can support. Accordingly, the default size of these fields has been reduced to 16-bits, which means you can have 32k procs in each of 32k jobs. Since the daemons must have a vpid, and we require one daemon/node, this also restricts the default configuration to 32k nodes.
To support any future "mega-clusters", a configuration option --enable-jumbo-apps has been added. This option increases the jobid and vpid field sizes to 32-bits. Someday, if necessary, someone can add yet another option to increase them to 64-bits, I suppose.
(b) replacing the string nodename with an integer nodeid. Since we have one daemon/node, the nodeid corresponds to the local daemon's vpid. This replaces an often lengthy string with only 2 (or at most 4) bytes, a substantial reduction.
(c) when the mca param requesting that nodenames be sent to support pretty error messages, a second mca param is now used to request FQDN - otherwise, the domain name is stripped (by default) from the message to save space. If someone wants to combine those into a single param somehow (perhaps with an argument?), they are welcome to do so - I didn't want to alter what people are already using.
While these may seem like small savings, they actually amount to a significant impact when aggregated across the entire modex operation. Since every proc must receive the modex data regardless of the collective used to send it, just reducing the size of the process name removes nearly 400MBytes of communication from a 32k proc job (admittedly, much of this comm may occur in parallel). So it does add up pretty quickly.
3. routing RML messages to reduce connections. The default messaging system remains point-to-point - i.e., each proc opens a socket to every proc it communicates with and sends its messages directly. A new option uses the orteds as routers - i.e., each proc only opens a single socket to its local orted. All messages are sent from the proc to the orted, which forwards the message to the orted on the node where the intended recipient proc is located - that orted then forwards the message to its local proc (the recipient). This greatly reduces the connection storm we have encountered during startup.
It also has the benefit of removing the sharing of every proc's OOB contact with every other proc. The orted routing tables are populated during launch since every orted gets a map of where every proc is being placed. Each proc, therefore, only needs to know the contact info for its local daemon, which is passed in via the environment when the proc is fork/exec'd by the daemon. This alone removes ~50 bytes/process of communication that was in the current STG1 startup message - so for our 32k proc job, this saves us roughly 32k*50 = 1.6MBytes sent to 32k procs = 51GBytes of messaging.
Note that you can use the new routing method by specifying -mca routed tree - if you so desire. This mode will become the default at some point in the future.
There are a few minor additional changes in the commit that I'll just note in passing:
* propagation of command line mca params to the orteds - fixes ticket #1073. See note there for details.
* requiring of "finalize" prior to "exit" for MPI procs - fixes ticket #1144. See note there for details.
* cleanup of some stale header files
This commit was SVN r16364.
2007-10-05 19:48:23 +00:00
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
/* prep the output buffer */
|
2008-05-05 22:32:25 +00:00
|
|
|
OBJ_CONSTRUCT(&buf, opal_buffer_t);
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_grpcomm_base_app_pack_xcast(ORTE_DAEMON_PROCESS_AND_RELAY_CMD,
|
|
|
|
|
job, &buf, buffer, tag))) {
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
/* if I am the HNP, just set things up so the cmd processor gets called.
|
|
|
|
|
* We don't want to message ourselves as this can create circular logic
|
|
|
|
|
* in the RML. Instead, this macro will set a zero-time event which will
|
|
|
|
|
* cause the buffer to be processed by the cmd processor - probably will
|
|
|
|
|
* fire right away, but that's okay
|
|
|
|
|
* The macro makes a copy of the buffer, so it's okay to release it here
|
|
|
|
|
*/
|
2009-05-04 11:07:40 +00:00
|
|
|
if (ORTE_PROC_IS_HNP) {
|
2008-05-05 22:32:25 +00:00
|
|
|
ORTE_MESSAGE_EVENT(ORTE_PROC_MY_NAME, &buf, ORTE_RML_TAG_DAEMON, orte_daemon_cmd_processor);
|
2008-04-09 22:10:53 +00:00
|
|
|
} else {
|
|
|
|
|
/* otherwise, send it to the HNP for relay */
|
2008-05-05 22:32:25 +00:00
|
|
|
if (0 > (rc = orte_rml.send_buffer(ORTE_PROC_MY_HNP, &buf, ORTE_RML_TAG_DAEMON, 0))) {
|
2008-03-03 20:07:02 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
rc = ORTE_SUCCESS;
|
|
|
|
|
}
|
2008-02-28 01:57:57 +00:00
|
|
|
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
CLEANUP:
|
2008-05-05 22:32:25 +00:00
|
|
|
OBJ_DESTRUCT(&buf);
|
Commit the orted-failed-to-start code. This correctly causes the system to detect the failure of an orted to start and allows the system to terminate all procs/orteds that *did* start.
The primary change that underlies all this is in the OOB. Specifically, the problem in the code until now has been that the OOB attempts to resolve an address when we call the "send" to an unknown recipient. The OOB would then wait forever if that recipient never actually started (and hence, never reported back its OOB contact info). In the case of an orted that failed to start, we would correctly detect that the orted hadn't started, but then we would attempt to order all orteds (including the one that failed to start) to die. This would cause the OOB to "hang" the system.
Unfortunately, revising how the OOB resolves addresses introduced a number of additional problems. Specifically, and most troublesome, was the fact that comm_spawn involved the immediate transmission of the rendezvous point from parent-to-child after the child was spawned. The current code used the OOB address resolution as a "barrier" - basically, the parent would attempt to send the info to the child, and then "hold" there until the child's contact info had arrived (meaning the child had started) and the send could be completed.
Note that this also caused comm_spawn to "hang" the entire system if the child never started... The app-failed-to-start helped improve that behavior - this code provides additional relief.
With this change, the OOB will return an ADDRESSEE_UNKNOWN error if you attempt to send to a recipient whose contact info isn't already in the OOB's hash tables. To resolve comm_spawn issues, we also now force the cross-sharing of connection info between parent and child jobs during spawn.
Finally, to aid in setting triggers to the right values, we introduce the "arith" API for the GPR. This function allows you to atomically change the value in a registry location (either divide, multiply, add, or subtract) by the provided operand. It is equivalent to first fetching the value using a "get", then modifying it, and then putting the result back into the registry via a "put".
This commit was SVN r14711.
2007-05-21 18:31:28 +00:00
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
|
|
|
|
|
static void barrier_recv(int status, orte_process_name_t* sender,
|
|
|
|
|
opal_buffer_t *buffer,
|
|
|
|
|
orte_rml_tag_t tag, void *cbdata)
|
|
|
|
|
{
|
2010-02-25 01:11:29 +00:00
|
|
|
orte_grpcomm_collective_t *coll = (orte_grpcomm_collective_t*)cbdata;
|
|
|
|
|
|
|
|
|
|
OPAL_THREAD_LOCK(&coll->lock);
|
2008-04-09 22:10:53 +00:00
|
|
|
/* flag as recvd */
|
2010-02-25 01:11:29 +00:00
|
|
|
coll->recvd = 1;
|
|
|
|
|
opal_condition_broadcast(&coll->cond);
|
|
|
|
|
OPAL_THREAD_UNLOCK(&coll->lock);
|
2008-04-09 22:10:53 +00:00
|
|
|
}
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
static int basic_barrier(void)
|
2008-04-09 22:10:53 +00:00
|
|
|
{
|
|
|
|
|
int rc;
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
|
2008-04-09 22:10:53 +00:00
|
|
|
"%s grpcomm:basic entering barrier",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
2008-04-09 22:10:53 +00:00
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
/* if I am alone, just return */
|
|
|
|
|
if (1 == orte_process_info.num_procs) {
|
|
|
|
|
return ORTE_SUCCESS;
|
2008-04-09 22:10:53 +00:00
|
|
|
}
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
/* setup the recv to get the response */
|
2008-04-09 22:10:53 +00:00
|
|
|
rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_BARRIER,
|
2010-02-25 01:11:29 +00:00
|
|
|
ORTE_RML_NON_PERSISTENT, barrier_recv, &barrier);
|
2008-04-09 22:10:53 +00:00
|
|
|
if (rc != ORTE_SUCCESS) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
/* send it and wait for the response */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_grpcomm_base_app_barrier(ORTE_PROC_MY_DAEMON, &barrier))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
2008-04-09 22:10:53 +00:00
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
/* don't need to cancel the recv as it only fires once */
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_grpcomm_base.output,
|
2008-04-09 22:10:53 +00:00
|
|
|
"%s grpcomm:basic received barrier release",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
2010-02-25 01:11:29 +00:00
|
|
|
|
|
|
|
|
return rc;
|
2009-05-11 14:11:44 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static void onesided_barrier_recv(int status, orte_process_name_t* sender,
|
|
|
|
|
opal_buffer_t* buffer, orte_rml_tag_t tag,
|
|
|
|
|
void* cbdata)
|
|
|
|
|
{
|
2010-02-25 01:11:29 +00:00
|
|
|
orte_grpcomm_collective_t *coll = (orte_grpcomm_collective_t*)cbdata;
|
2009-05-11 14:11:44 +00:00
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_THREAD_LOCK(&coll->lock);
|
|
|
|
|
/* flag as recvd */
|
|
|
|
|
coll->recvd += 1;
|
|
|
|
|
if (orte_process_info.num_procs == coll->recvd) {
|
|
|
|
|
opal_condition_broadcast(&coll->cond);
|
2009-05-11 14:11:44 +00:00
|
|
|
}
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_THREAD_UNLOCK(&coll->lock);
|
2009-05-11 14:11:44 +00:00
|
|
|
}
|
|
|
|
|
/* quick timeout loop */
|
|
|
|
|
static bool timer_fired;
|
|
|
|
|
|
|
|
|
|
static void quicktime_cb(int fd, short event, void *cbdata)
|
|
|
|
|
{
|
|
|
|
|
/* declare it fired */
|
|
|
|
|
timer_fired = true;
|
|
|
|
|
}
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
static int basic_onesided_barrier(void)
|
2009-05-11 14:11:44 +00:00
|
|
|
{
|
|
|
|
|
opal_list_t daemon_tree;
|
2010-02-25 01:11:29 +00:00
|
|
|
opal_list_item_t *item;
|
2009-05-11 14:11:44 +00:00
|
|
|
opal_buffer_t buf;
|
|
|
|
|
orte_process_name_t my_parent;
|
|
|
|
|
opal_event_t *quicktime=NULL;
|
|
|
|
|
struct timeval quicktimeval;
|
|
|
|
|
int rc;
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((5, orte_grpcomm_base.output,
|
2009-05-11 14:11:44 +00:00
|
|
|
"%s grpcomm:basic: onesided barrier called",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
|
|
|
|
/* if we are not to use the barrier, then just return */
|
|
|
|
|
if (!orte_orted_exit_with_barrier) {
|
|
|
|
|
if (ORTE_PROC_IS_HNP) {
|
|
|
|
|
/* if we are the HNP, we need to do a little delay to give
|
|
|
|
|
* the orteds a chance to exit before we leave
|
|
|
|
|
*/
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((5, orte_grpcomm_base.output,
|
2009-05-11 14:11:44 +00:00
|
|
|
"%s grpcomm:basic: onesided barrier adding delay timer",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
quicktimeval.tv_sec = 0;
|
|
|
|
|
quicktimeval.tv_usec = 100;
|
|
|
|
|
timer_fired = false;
|
|
|
|
|
ORTE_DETECT_TIMEOUT(&quicktime, orte_process_info.num_procs, 1000, 10000, quicktime_cb);
|
|
|
|
|
ORTE_PROGRESSED_WAIT(timer_fired, 0, 1);
|
|
|
|
|
}
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* figure out how many participants we should be expecting */
|
|
|
|
|
OBJ_CONSTRUCT(&daemon_tree, opal_list_t);
|
|
|
|
|
my_parent.jobid = ORTE_PROC_MY_NAME->jobid;
|
|
|
|
|
my_parent.vpid = orte_routed.get_routing_tree(&daemon_tree);
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_THREAD_LOCK(&onesided_barrier.lock);
|
|
|
|
|
onesided_barrier.recvd += orte_process_info.num_procs - opal_list_get_size(&daemon_tree);
|
|
|
|
|
OPAL_THREAD_UNLOCK(&onesided_barrier.lock);
|
2009-05-11 14:11:44 +00:00
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((5, orte_grpcomm_base.output,
|
2009-05-11 14:11:44 +00:00
|
|
|
"%s grpcomm:basic: onesided barrier num_participating %d",
|
2010-02-25 01:11:29 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
(int)(orte_process_info.num_procs - opal_list_get_size(&daemon_tree))));
|
|
|
|
|
|
|
|
|
|
/* disassemble the daemon tree */
|
|
|
|
|
while (NULL != (item = opal_list_remove_first(&daemon_tree))) {
|
|
|
|
|
OBJ_RELEASE(item);
|
|
|
|
|
}
|
|
|
|
|
OBJ_DESTRUCT(&daemon_tree);
|
2009-05-11 14:11:44 +00:00
|
|
|
|
|
|
|
|
/* set the recv */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD,
|
|
|
|
|
ORTE_RML_TAG_ONESIDED_BARRIER,
|
2010-02-25 01:11:29 +00:00
|
|
|
ORTE_RML_PERSISTENT,
|
2009-05-11 14:11:44 +00:00
|
|
|
onesided_barrier_recv,
|
2010-02-25 01:11:29 +00:00
|
|
|
&onesided_barrier))) {
|
2009-05-11 14:11:44 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
/* wait to get all my inputs */
|
|
|
|
|
OPAL_THREAD_LOCK(&onesided_barrier.lock);
|
|
|
|
|
while (onesided_barrier.recvd < orte_process_info.num_procs) {
|
|
|
|
|
opal_condition_wait(&onesided_barrier.cond, &onesided_barrier.lock);
|
|
|
|
|
}
|
|
|
|
|
/* reset the collective */
|
|
|
|
|
onesided_barrier.recvd = 0;
|
|
|
|
|
OPAL_THREAD_UNLOCK(&onesided_barrier.lock);
|
2009-05-11 14:11:44 +00:00
|
|
|
|
|
|
|
|
/* cancel the recv */
|
|
|
|
|
orte_rml.recv_cancel(ORTE_NAME_WILDCARD, ORTE_RML_TAG_ONESIDED_BARRIER);
|
|
|
|
|
|
|
|
|
|
/* if I am the HNP, then we are done */
|
|
|
|
|
if (ORTE_PROC_IS_HNP) {
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* send a zero-byte msg to my parent */
|
|
|
|
|
OBJ_CONSTRUCT(&buf, opal_buffer_t);
|
|
|
|
|
/* send it */
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((5, orte_grpcomm_base.output,
|
2009-05-11 14:11:44 +00:00
|
|
|
"%s grpcomm:basic:onsided:barrier not the HNP - sending to parent %s",
|
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_NAME_PRINT(&my_parent)));
|
|
|
|
|
if (0 > (rc = orte_rml.send_buffer(&my_parent, &buf, ORTE_RML_TAG_ONESIDED_BARRIER, 0))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
OBJ_DESTRUCT(&buf);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
OBJ_DESTRUCT(&buf);
|
|
|
|
|
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2008-04-09 22:10:53 +00:00
|
|
|
static void allgather_recv(int status, orte_process_name_t* sender,
|
2010-02-25 01:11:29 +00:00
|
|
|
opal_buffer_t *buffer,
|
|
|
|
|
orte_rml_tag_t tag, void *cbdata)
|
2008-03-03 20:07:02 +00:00
|
|
|
{
|
2010-02-25 01:11:29 +00:00
|
|
|
orte_grpcomm_collective_t *coll = (orte_grpcomm_collective_t*)cbdata;
|
2008-04-09 22:10:53 +00:00
|
|
|
int rc;
|
2008-03-03 20:07:02 +00:00
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_THREAD_LOCK(&coll->lock);
|
2008-04-09 22:10:53 +00:00
|
|
|
/* xfer the data */
|
2010-02-25 01:11:29 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.copy_payload(&coll->results, buffer))) {
|
2008-04-09 22:10:53 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
2010-02-25 01:11:29 +00:00
|
|
|
/* the daemon returns ALL of our recipients in a single message */
|
|
|
|
|
coll->recvd = orte_process_info.num_procs;
|
|
|
|
|
opal_condition_broadcast(&coll->cond);
|
|
|
|
|
OPAL_THREAD_UNLOCK(&coll->lock);
|
2008-04-09 22:10:53 +00:00
|
|
|
}
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
static int basic_allgather(opal_buffer_t *sbuf, opal_buffer_t *rbuf)
|
2008-04-09 22:10:53 +00:00
|
|
|
{
|
|
|
|
|
int rc;
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
|
2008-04-09 22:10:53 +00:00
|
|
|
"%s grpcomm:basic entering allgather",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
2008-04-09 22:10:53 +00:00
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
/* setup to receive results */
|
2008-04-09 22:10:53 +00:00
|
|
|
rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD, ORTE_RML_TAG_ALLGATHER,
|
2010-02-25 01:11:29 +00:00
|
|
|
ORTE_RML_NON_PERSISTENT, allgather_recv, &allgather);
|
2008-04-09 22:10:53 +00:00
|
|
|
if (rc != ORTE_SUCCESS) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
2008-03-03 20:07:02 +00:00
|
|
|
}
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
/* everyone sends data to their local daemon */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_grpcomm_base_app_allgather(ORTE_PROC_MY_DAEMON,
|
|
|
|
|
&allgather, sbuf, rbuf))) {
|
2008-04-09 22:10:53 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
/* don't need to cancel the recv as it only fires once */
|
|
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
|
2008-04-09 22:10:53 +00:00
|
|
|
"%s grpcomm:basic allgather completed",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
2010-02-25 01:11:29 +00:00
|
|
|
|
2008-03-03 20:07:02 +00:00
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
2008-06-18 22:17:53 +00:00
|
|
|
|
|
|
|
|
/*** MODEX SECTION ***/
|
2009-01-07 15:00:26 +00:00
|
|
|
static int modex(opal_list_t *procs)
|
|
|
|
|
{
|
|
|
|
|
int rc=ORTE_SUCCESS;
|
|
|
|
|
int fd;
|
|
|
|
|
opal_byte_object_t bo, *boptr;
|
|
|
|
|
int32_t i, n;
|
|
|
|
|
char *nodename, *attr;
|
|
|
|
|
orte_nid_t **nd, *ndptr;
|
|
|
|
|
orte_attr_t *attrdata;
|
|
|
|
|
opal_buffer_t bobuf;
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
|
2008-06-18 22:17:53 +00:00
|
|
|
"%s grpcomm:basic: modex entered",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
2008-06-18 22:17:53 +00:00
|
|
|
|
2008-12-09 23:49:02 +00:00
|
|
|
/* if we were given a list of procs to modex with, then this is happening
|
|
|
|
|
* as part of a connect/accept operation. In this case, we -must- do the
|
|
|
|
|
* modex for two reasons:
|
|
|
|
|
*
|
|
|
|
|
* (a) the modex could involve procs from different mpiruns. In this case,
|
|
|
|
|
* there is no way for the two sets of procs to know which node the
|
|
|
|
|
* other procs are on, so we cannot use the profile_file to determine
|
|
|
|
|
* their contact info
|
|
|
|
|
*
|
|
|
|
|
* (b) in a comm_spawn, the parent job does not have a pidmap for the
|
|
|
|
|
* child job. Thus, it cannot know where the child procs are located,
|
|
|
|
|
* and cannot use the profile_file to determine their contact info
|
2009-01-07 15:00:26 +00:00
|
|
|
*
|
2008-06-19 13:48:26 +00:00
|
|
|
*/
|
2009-02-06 15:25:06 +00:00
|
|
|
if (NULL != procs) {
|
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_grpcomm_base_full_modex(procs, false))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* Do a modex across our peers if we are doing an opal_profile so that the
|
|
|
|
|
* HNP can collect our modex info
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
if (opal_profile) {
|
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_grpcomm_base_peer_modex(false))) {
|
2009-01-07 15:00:26 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
|
|
|
|
return rc;
|
2009-02-06 15:25:06 +00:00
|
|
|
}
|
|
|
|
|
|
2009-05-06 20:11:28 +00:00
|
|
|
if (OPAL_ENABLE_HETEROGENEOUS_SUPPORT) {
|
2008-12-09 23:49:02 +00:00
|
|
|
/* decide if we need to add the architecture to the modex. Check
|
|
|
|
|
* first to see if hetero is enabled - if not, then we clearly
|
|
|
|
|
* don't need to exchange arch's as they are all identical
|
|
|
|
|
*/
|
2008-07-01 02:44:57 +00:00
|
|
|
/* Case 1: If different apps in this job were built differently - e.g., some
|
|
|
|
|
* are built 32-bit while others are built 64-bit - then we need to modex
|
|
|
|
|
* regardless of any other consideration. The user is reqd to tell us via a
|
|
|
|
|
* cmd line option if this situation exists, which will result in an mca param
|
|
|
|
|
* being set for us, so all we need to do is check for the global boolean
|
|
|
|
|
* that corresponds to that param
|
2008-11-03 21:48:52 +00:00
|
|
|
*
|
|
|
|
|
* Case 2: the nodes are hetero, but the app binaries were built
|
2008-07-01 02:44:57 +00:00
|
|
|
* the same - i.e., either they are both 32-bit, or they are both 64-bit, but
|
|
|
|
|
* no mixing of the two. In this case, we include the info in the modex
|
|
|
|
|
*/
|
2008-11-03 21:48:52 +00:00
|
|
|
if (orte_hetero_apps || !orte_homogeneous_nodes) {
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
|
2008-11-03 21:48:52 +00:00
|
|
|
"%s grpcomm:basic: modex is required",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
2008-11-03 21:48:52 +00:00
|
|
|
|
2009-02-06 15:25:06 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_grpcomm_base_peer_modex(false))) {
|
2009-01-07 15:00:26 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
|
|
|
|
return rc;
|
2008-07-01 02:44:57 +00:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
2009-01-07 15:00:26 +00:00
|
|
|
/* no modex is required - see if the data was included in the launch message */
|
|
|
|
|
if (orte_send_profile) {
|
|
|
|
|
/* the info was provided in the nidmap - there is nothing more we have to do */
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
|
2009-01-07 15:00:26 +00:00
|
|
|
"%s grpcomm:basic:modex using nidmap",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
2009-01-07 15:00:26 +00:00
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* see if a profile file was given to us */
|
|
|
|
|
if (NULL == opal_profile_file) {
|
|
|
|
|
/* if we don't have any other way to do this, then let's default to doing the
|
|
|
|
|
* modex so we at least can function, even if it isn't as fast as we might like
|
|
|
|
|
*/
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
|
2009-01-07 15:00:26 +00:00
|
|
|
"%s grpcomm:basic: modex is required",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
2009-02-06 15:25:06 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = orte_grpcomm_base_peer_modex(false))) {
|
2008-12-09 23:49:02 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
2009-01-07 15:00:26 +00:00
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
fd = open(opal_profile_file, O_RDONLY);
|
|
|
|
|
if (fd < 0) {
|
|
|
|
|
orte_show_help("help-orte-runtime.txt", "grpcomm-basic:file-cant-open", true, opal_profile_file);
|
|
|
|
|
return ORTE_ERR_NOT_FOUND;
|
|
|
|
|
}
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
|
2009-01-07 15:00:26 +00:00
|
|
|
"%s grpcomm:basic:modex reading %s file",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), opal_profile_file));
|
2009-01-07 15:00:26 +00:00
|
|
|
|
|
|
|
|
/* loop through file until end */
|
|
|
|
|
boptr = &bo;
|
|
|
|
|
nd = (orte_nid_t**)orte_nidmap.addr;
|
|
|
|
|
while (0 < read(fd, &bo.size, sizeof(bo.size))) {
|
|
|
|
|
/* this is the number of bytes in the byte object */
|
2009-02-05 16:37:44 +00:00
|
|
|
bo.bytes = (uint8_t *) malloc(bo.size);
|
2009-01-07 15:00:26 +00:00
|
|
|
if (0 > read(fd, bo.bytes, bo.size)) {
|
|
|
|
|
orte_show_help("help-orte-runtime.txt", "orte_nidmap:unable-read-file", true, opal_profile_file);
|
|
|
|
|
close(fd);
|
|
|
|
|
return ORTE_ERR_FILE_READ_FAILURE;
|
|
|
|
|
}
|
|
|
|
|
/* load the byte object into a buffer for unpacking */
|
|
|
|
|
OBJ_CONSTRUCT(&bobuf, opal_buffer_t);
|
|
|
|
|
opal_dss.load(&bobuf, boptr->bytes, boptr->size);
|
|
|
|
|
/* unpack the nodename */
|
|
|
|
|
n = 1;
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.unpack(&bobuf, &nodename, &n, OPAL_STRING))) {
|
2008-12-09 23:49:02 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
2009-01-07 15:00:26 +00:00
|
|
|
return rc;
|
2008-12-09 23:49:02 +00:00
|
|
|
}
|
2009-01-07 15:00:26 +00:00
|
|
|
/* find this node in nidmap */
|
|
|
|
|
for (i=0, ndptr=NULL; i < orte_nidmap.size && NULL != nd[i]; i++) {
|
|
|
|
|
/* since we may not have kept fqdn hostnames, we can only check
|
|
|
|
|
* for equality to the length of the name in the nid
|
|
|
|
|
*/
|
|
|
|
|
if (0 == strncmp(nd[i]->name, nodename, strlen(nd[i]->name))) {
|
|
|
|
|
ndptr = nd[i];
|
|
|
|
|
break;
|
2008-06-18 22:17:53 +00:00
|
|
|
}
|
|
|
|
|
}
|
2009-01-07 15:00:26 +00:00
|
|
|
free(nodename); /* done with this */
|
|
|
|
|
if (NULL == ndptr) {
|
|
|
|
|
/* didn't find it! */
|
|
|
|
|
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
|
|
|
|
|
return ORTE_ERR_NOT_FOUND;
|
2008-06-18 22:17:53 +00:00
|
|
|
}
|
|
|
|
|
|
2009-01-07 15:00:26 +00:00
|
|
|
/* loop through the rest of the object to unpack the attr's themselves */
|
|
|
|
|
n = 1;
|
|
|
|
|
while (ORTE_SUCCESS == opal_dss.unpack(&bobuf, &attr, &n, OPAL_STRING)) {
|
|
|
|
|
attrdata = OBJ_NEW(orte_attr_t);
|
|
|
|
|
attrdata->name = strdup(attr);
|
|
|
|
|
/* read the number of bytes in the blob */
|
|
|
|
|
n = 1;
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.unpack(&bobuf, &attrdata->size, &n, OPAL_INT32))) {
|
2008-06-18 22:17:53 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
2009-01-07 15:00:26 +00:00
|
|
|
return rc;
|
2008-06-18 22:17:53 +00:00
|
|
|
}
|
2009-01-07 15:00:26 +00:00
|
|
|
/* unpack the bytes */
|
2009-02-05 16:37:44 +00:00
|
|
|
attrdata->bytes = (uint8_t *) malloc(attrdata->size);
|
2009-01-07 15:00:26 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.unpack(&bobuf, attrdata->bytes, &attrdata->size, OPAL_BYTE))) {
|
2008-06-18 22:17:53 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
2009-01-07 15:00:26 +00:00
|
|
|
return rc;
|
2008-06-18 22:17:53 +00:00
|
|
|
}
|
2009-01-07 15:00:26 +00:00
|
|
|
/* add to our list for this node */
|
|
|
|
|
opal_list_append(&ndptr->attrs, &attrdata->super);
|
2008-06-18 22:17:53 +00:00
|
|
|
}
|
2009-01-07 15:00:26 +00:00
|
|
|
OBJ_DESTRUCT(&bobuf);
|
2008-06-18 22:17:53 +00:00
|
|
|
}
|
2009-01-07 15:00:26 +00:00
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
|
2008-06-18 22:17:53 +00:00
|
|
|
"%s grpcomm:basic: modex completed",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
2008-06-18 22:17:53 +00:00
|
|
|
|
|
|
|
|
return rc;
|
|
|
|
|
}
|
|
|
|
|
|
2008-12-09 23:49:02 +00:00
|
|
|
/* the HNP will -never- execute the following as it is NOT an MPI process */
|
|
|
|
|
static int set_proc_attr(const char *attr_name, const void *data, size_t size)
|
|
|
|
|
{
|
|
|
|
|
int rc;
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_grpcomm_base.output,
|
2008-12-09 23:49:02 +00:00
|
|
|
"%s grpcomm:basic:set_proc_attr for attribute %s",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), attr_name));
|
2008-12-09 23:49:02 +00:00
|
|
|
|
2009-01-07 15:00:26 +00:00
|
|
|
/* if we are doing a profile, pack this up */
|
2008-12-09 23:49:02 +00:00
|
|
|
if (opal_profile) {
|
|
|
|
|
int32_t isize;
|
|
|
|
|
|
2009-01-07 15:00:26 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(profile_buf, &attr_name, 1, OPAL_STRING))) {
|
2008-12-09 23:49:02 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto cleanup;
|
|
|
|
|
}
|
|
|
|
|
isize = size;
|
2009-01-07 15:00:26 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(profile_buf, &isize, 1, OPAL_INT32))) {
|
2008-12-09 23:49:02 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto cleanup;
|
|
|
|
|
}
|
2009-01-07 15:00:26 +00:00
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.pack(profile_buf, data, isize, OPAL_BYTE))) {
|
2008-12-09 23:49:02 +00:00
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto cleanup;
|
|
|
|
|
}
|
|
|
|
|
/* let it fall through so that the job doesn't hang! */
|
|
|
|
|
return orte_grpcomm_base_set_proc_attr(attr_name, data, size);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/* we always have to set our own attributes in case they are needed for
|
|
|
|
|
* a connect/accept at some later time
|
|
|
|
|
*/
|
2009-01-07 15:00:26 +00:00
|
|
|
cleanup:
|
|
|
|
|
return orte_grpcomm_base_set_proc_attr(attr_name, data, size);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
static int get_proc_attr(const orte_process_name_t proc,
|
|
|
|
|
const char * attribute_name, void **val,
|
|
|
|
|
size_t *size)
|
|
|
|
|
{
|
|
|
|
|
orte_nid_t *nid;
|
|
|
|
|
opal_list_item_t *item;
|
|
|
|
|
orte_attr_t *attr;
|
2008-12-09 23:49:02 +00:00
|
|
|
|
2009-01-07 15:00:26 +00:00
|
|
|
/* find this proc's node in the nidmap */
|
|
|
|
|
if (NULL == (nid = orte_util_lookup_nid((orte_process_name_t*)&proc))) {
|
|
|
|
|
/* proc wasn't found - return error */
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((5, orte_grpcomm_base.output,
|
2009-01-07 15:00:26 +00:00
|
|
|
"%s grpcomm:basic:get_proc_attr: no modex entry for proc %s",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_NAME_PRINT(&proc)));
|
2008-12-09 23:49:02 +00:00
|
|
|
return ORTE_ERR_NOT_FOUND;
|
2009-01-07 15:00:26 +00:00
|
|
|
|
2008-12-09 23:49:02 +00:00
|
|
|
}
|
|
|
|
|
|
2009-01-07 15:00:26 +00:00
|
|
|
/* look for this attribute */
|
|
|
|
|
for (item = opal_list_get_first(&nid->attrs);
|
|
|
|
|
item != opal_list_get_end(&nid->attrs);
|
|
|
|
|
item = opal_list_get_next(item)) {
|
|
|
|
|
attr = (orte_attr_t*)item;
|
|
|
|
|
if (0 == strcmp(attr->name, attribute_name)) {
|
|
|
|
|
/* copy the data to the caller */
|
|
|
|
|
void *copy = malloc(attr->size);
|
|
|
|
|
|
|
|
|
|
if (copy == NULL) {
|
|
|
|
|
return ORTE_ERR_OUT_OF_RESOURCE;
|
2008-12-09 23:49:02 +00:00
|
|
|
}
|
2009-01-07 15:00:26 +00:00
|
|
|
memcpy(copy, attr->bytes, attr->size);
|
|
|
|
|
*val = copy;
|
|
|
|
|
*size = attr->size;
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((5, orte_grpcomm_base.output,
|
2009-01-07 15:00:26 +00:00
|
|
|
"%s grpcomm:basic:get_proc_attr: found %d bytes for attr %s on proc %s",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), (int)attr->size,
|
|
|
|
|
attribute_name, ORTE_NAME_PRINT(&proc)));
|
2009-01-07 15:00:26 +00:00
|
|
|
return ORTE_SUCCESS;
|
2008-12-09 23:49:02 +00:00
|
|
|
}
|
|
|
|
|
}
|
2009-01-07 15:00:26 +00:00
|
|
|
|
|
|
|
|
/* get here if attribute isn't found */
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((5, orte_grpcomm_base.output,
|
2009-01-07 15:00:26 +00:00
|
|
|
"%s grpcomm:basic:get_proc_attr: no attr avail or zero byte size for proc %s attribute %s",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_NAME_PRINT(&proc), attribute_name));
|
2009-01-07 15:00:26 +00:00
|
|
|
*val = NULL;
|
|
|
|
|
*size = 0;
|
|
|
|
|
|
2008-12-09 23:49:02 +00:00
|
|
|
return ORTE_SUCCESS;
|
|
|
|
|
}
|
2009-01-27 19:13:56 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
/* process incoming messages in order of receipt */
|
|
|
|
|
static void process_msg(int fd, short event, void *data)
|
|
|
|
|
{
|
|
|
|
|
orte_message_event_t *mev = (orte_message_event_t*)data;
|
|
|
|
|
int32_t rc, count;
|
|
|
|
|
opal_byte_object_t *bo;
|
|
|
|
|
|
|
|
|
|
/* save the info in the file */
|
|
|
|
|
if (0 <= profile_fd) {
|
|
|
|
|
/* extract the byte object holding the node's modex info */
|
|
|
|
|
count=1;
|
|
|
|
|
if (ORTE_SUCCESS != (rc = opal_dss.unpack(mev->buffer, &bo, &count, OPAL_BYTE_OBJECT))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
goto CLEANUP;
|
|
|
|
|
}
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((5, orte_grpcomm_base.output,
|
2009-01-27 19:13:56 +00:00
|
|
|
"%s grpcomm:basic:receive:profile writing %d bytes of data from proc %s",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
bo->size, ORTE_NAME_PRINT(&mev->sender)));
|
2009-01-27 19:13:56 +00:00
|
|
|
|
|
|
|
|
write(profile_fd, &bo->size, sizeof(bo->size));
|
|
|
|
|
write(profile_fd, bo->bytes, bo->size);
|
|
|
|
|
free(bo->bytes);
|
|
|
|
|
free(bo);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
CLEANUP:
|
|
|
|
|
/* release the message */
|
|
|
|
|
OBJ_RELEASE(mev);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
/*
|
|
|
|
|
* NOTE: The incoming buffer "buffer" is OBJ_RELEASED by the calling program.
|
|
|
|
|
* DO NOT RELEASE THIS BUFFER IN THIS CODE
|
|
|
|
|
*/
|
|
|
|
|
|
|
|
|
|
static void profile_recv(int status, orte_process_name_t* sender,
|
|
|
|
|
opal_buffer_t* buffer, orte_rml_tag_t tag,
|
|
|
|
|
void* cbdata)
|
|
|
|
|
{
|
|
|
|
|
int rc;
|
|
|
|
|
|
2010-02-25 01:11:29 +00:00
|
|
|
OPAL_OUTPUT_VERBOSE((5, orte_grpcomm_base.output,
|
2009-01-27 19:13:56 +00:00
|
|
|
"%s grpcomm:basic:receive got message from %s",
|
2009-03-05 21:50:47 +00:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
|
ORTE_NAME_PRINT(sender)));
|
2009-01-27 19:13:56 +00:00
|
|
|
|
|
|
|
|
/* 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, process_msg);
|
|
|
|
|
|
|
|
|
|
/* reissue the recv */
|
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml.recv_buffer_nb(ORTE_NAME_WILDCARD,
|
|
|
|
|
ORTE_RML_TAG_GRPCOMM_PROFILE,
|
|
|
|
|
ORTE_RML_NON_PERSISTENT,
|
|
|
|
|
profile_recv,
|
|
|
|
|
NULL))) {
|
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
|
}
|
|
|
|
|
return;
|
|
|
|
|
}
|