2008-05-06 02:32:25 +04:00
|
|
|
/*
|
|
|
|
* Copyright (c) 2007 Los Alamos National Security, LLC.
|
|
|
|
* All rights reserved.
|
|
|
|
* $COPYRIGHT$
|
|
|
|
*
|
|
|
|
* Additional copyrights may follow
|
|
|
|
*
|
|
|
|
* $HEADER$
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include "orte_config.h"
|
|
|
|
#include "orte/constants.h"
|
|
|
|
|
|
|
|
#include "opal/threads/condition.h"
|
|
|
|
#include "opal/runtime/opal_progress.h"
|
|
|
|
#include "opal/dss/dss.h"
|
|
|
|
#include "opal/class/opal_hash_table.h"
|
2008-10-03 00:08:27 +04:00
|
|
|
#include "opal/class/opal_bitmap.h"
|
2008-05-06 02:32:25 +04:00
|
|
|
#include "opal/util/bit_ops.h"
|
2009-02-14 05:26:12 +03:00
|
|
|
#include "opal/util/output.h"
|
2008-05-06 02:32:25 +04:00
|
|
|
|
|
|
|
#include "orte/mca/errmgr/errmgr.h"
|
2008-11-01 00:10:00 +03:00
|
|
|
#include "orte/mca/ess/ess.h"
|
2008-05-06 02:32:25 +04:00
|
|
|
#include "orte/mca/grpcomm/grpcomm.h"
|
|
|
|
#include "orte/mca/rml/rml.h"
|
2009-02-14 05:26:12 +03:00
|
|
|
#include "orte/mca/rml/rml_types.h"
|
2008-05-06 02:32:25 +04:00
|
|
|
#include "orte/mca/odls/odls_types.h"
|
|
|
|
#include "orte/util/name_fns.h"
|
|
|
|
#include "orte/runtime/orte_globals.h"
|
|
|
|
#include "orte/runtime/orte_wait.h"
|
2008-06-18 07:15:56 +04:00
|
|
|
#include "orte/runtime/runtime.h"
|
2008-05-06 02:32:25 +04:00
|
|
|
|
|
|
|
#include "orte/mca/rml/base/rml_contact.h"
|
|
|
|
|
|
|
|
#include "orte/mca/routed/base/base.h"
|
|
|
|
#include "routed_binomial.h"
|
|
|
|
|
|
|
|
static int init(void);
|
|
|
|
static int finalize(void);
|
2008-07-22 01:37:09 +04:00
|
|
|
static int delete_route(orte_process_name_t *proc);
|
2008-05-06 02:32:25 +04:00
|
|
|
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);
|
|
|
|
static bool route_is_defined(const orte_process_name_t *target);
|
|
|
|
static int update_routing_tree(void);
|
2009-01-27 22:13:56 +03:00
|
|
|
static orte_vpid_t get_routing_tree(opal_list_t *children);
|
2008-10-03 00:08:27 +04:00
|
|
|
static int get_wireup_info(opal_buffer_t *buf);
|
2008-12-21 02:23:11 +03:00
|
|
|
static int set_lifeline(orte_process_name_t *proc);
|
2008-05-06 02:32:25 +04:00
|
|
|
|
|
|
|
#if OPAL_ENABLE_FT == 1
|
|
|
|
static int binomial_ft_event(int state);
|
|
|
|
#endif
|
|
|
|
|
|
|
|
orte_routed_module_t orte_routed_binomial_module = {
|
|
|
|
init,
|
|
|
|
finalize,
|
2008-07-22 01:37:09 +04:00
|
|
|
delete_route,
|
2008-05-06 02:32:25 +04:00
|
|
|
update_route,
|
|
|
|
get_route,
|
|
|
|
init_routes,
|
|
|
|
route_lost,
|
|
|
|
route_is_defined,
|
2008-12-21 02:23:11 +03:00
|
|
|
set_lifeline,
|
2008-05-06 02:32:25 +04:00
|
|
|
update_routing_tree,
|
|
|
|
get_routing_tree,
|
|
|
|
get_wireup_info,
|
|
|
|
#if OPAL_ENABLE_FT == 1
|
|
|
|
binomial_ft_event
|
|
|
|
#else
|
|
|
|
NULL
|
|
|
|
#endif
|
|
|
|
};
|
|
|
|
|
|
|
|
/* local globals */
|
2008-11-01 00:10:00 +03:00
|
|
|
static opal_hash_table_t jobfam_list;
|
2008-05-06 02:32:25 +04:00
|
|
|
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 21:53:37 +04:00
|
|
|
static orte_process_name_t *lifeline=NULL;
|
2008-12-21 02:23:11 +03:00
|
|
|
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 21:53:37 +04:00
|
|
|
static orte_process_name_t my_parent;
|
|
|
|
static int num_children;
|
|
|
|
static opal_list_t my_children;
|
|
|
|
static bool ack_recvd;
|
2008-05-06 02:32:25 +04:00
|
|
|
|
|
|
|
|
|
|
|
static int init(void)
|
|
|
|
{
|
2008-11-01 00:10:00 +03:00
|
|
|
OBJ_CONSTRUCT(&jobfam_list, opal_hash_table_t);
|
|
|
|
opal_hash_table_init(&jobfam_list, 128);
|
2008-05-06 02:32:25 +04:00
|
|
|
|
|
|
|
/* setup the global condition and lock */
|
|
|
|
OBJ_CONSTRUCT(&cond, opal_condition_t);
|
|
|
|
OBJ_CONSTRUCT(&lock, opal_mutex_t);
|
|
|
|
|
|
|
|
lifeline = NULL;
|
|
|
|
|
|
|
|
/* setup the list of children */
|
|
|
|
OBJ_CONSTRUCT(&my_children, opal_list_t);
|
|
|
|
num_children = 0;
|
|
|
|
my_parent.jobid = ORTE_PROC_MY_NAME->jobid;
|
|
|
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int finalize(void)
|
|
|
|
{
|
|
|
|
int rc;
|
|
|
|
opal_list_item_t *item;
|
|
|
|
|
|
|
|
/* if I am an application process, indicate that I am
|
|
|
|
* truly finalizing prior to departure
|
|
|
|
*/
|
|
|
|
if (!orte_process_info.hnp &&
|
|
|
|
!orte_process_info.daemon &&
|
|
|
|
!orte_process_info.tool) {
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_routed_base_register_sync(false))) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* if I am the HNP, I need to stop the comm recv */
|
|
|
|
if (orte_process_info.hnp) {
|
|
|
|
orte_routed_base_comm_stop();
|
|
|
|
}
|
|
|
|
|
2008-11-01 00:10:00 +03:00
|
|
|
OBJ_DESTRUCT(&jobfam_list);
|
2008-05-06 02:32:25 +04:00
|
|
|
/* destruct the global condition and lock */
|
|
|
|
OBJ_DESTRUCT(&cond);
|
|
|
|
OBJ_DESTRUCT(&lock);
|
|
|
|
|
|
|
|
lifeline = NULL;
|
|
|
|
|
|
|
|
/* deconstruct the list of children */
|
|
|
|
while (NULL != (item = opal_list_remove_first(&my_children))) {
|
|
|
|
OBJ_RELEASE(item);
|
|
|
|
}
|
|
|
|
OBJ_DESTRUCT(&my_children);
|
|
|
|
num_children = 0;
|
|
|
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
}
|
|
|
|
|
2008-07-22 01:37:09 +04:00
|
|
|
static int delete_route(orte_process_name_t *proc)
|
|
|
|
{
|
|
|
|
int rc;
|
|
|
|
orte_process_name_t *route_copy;
|
|
|
|
|
|
|
|
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
|
|
|
|
*/
|
|
|
|
if (!orte_process_info.hnp && !orte_process_info.daemon &&
|
|
|
|
!orte_process_info.tool) {
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
}
|
|
|
|
|
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
|
|
|
"%s routed_binomial_delete_route for %s",
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
ORTE_NAME_PRINT(proc)));
|
|
|
|
|
|
|
|
|
|
|
|
/* 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
|
|
|
|
*/
|
|
|
|
if (orte_process_info.daemon) {
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* see if this proc is present - it will have a wildcard vpid,
|
|
|
|
* so we have to look for it with that condition
|
|
|
|
*/
|
2008-11-01 00:10:00 +03:00
|
|
|
rc = opal_hash_table_get_value_uint32(&jobfam_list,
|
2008-07-22 01:37:09 +04:00
|
|
|
ORTE_JOB_FAMILY(proc->jobid),
|
|
|
|
(void**)&route_copy);
|
|
|
|
if (ORTE_SUCCESS == rc && NULL != route_copy) {
|
|
|
|
/* proc is present - remove the data */
|
|
|
|
free(route_copy);
|
2008-11-01 00:10:00 +03:00
|
|
|
rc = opal_hash_table_remove_value_uint32(&jobfam_list,
|
2008-07-22 01:37:09 +04:00
|
|
|
ORTE_JOB_FAMILY(proc->jobid));
|
|
|
|
if (ORTE_SUCCESS != rc) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
}
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* not present - nothing to do */
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
}
|
|
|
|
|
2008-11-01 00:10:00 +03:00
|
|
|
/* THIS CAME FROM OUR OWN JOB FAMILY...there is nothing
|
|
|
|
* to do here. The routes will be redefined when we update
|
|
|
|
* the routing tree
|
2008-07-22 01:37:09 +04:00
|
|
|
*/
|
2008-11-01 00:10:00 +03:00
|
|
|
|
|
|
|
return ORTE_SUCCESS;
|
2008-07-22 01:37:09 +04:00
|
|
|
}
|
|
|
|
|
2008-05-06 02:32:25 +04:00
|
|
|
static int update_route(orte_process_name_t *target,
|
|
|
|
orte_process_name_t *route)
|
|
|
|
{
|
|
|
|
int rc;
|
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 21:53:37 +04:00
|
|
|
orte_process_name_t *route_copy;
|
2008-05-06 02:32:25 +04:00
|
|
|
|
|
|
|
if (target->jobid == ORTE_JOBID_INVALID ||
|
|
|
|
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 21:53:37 +04:00
|
|
|
/* if I am an application process, we don't update the route since
|
|
|
|
* we automatically route everything through the local daemon
|
|
|
|
*/
|
|
|
|
if (!orte_process_info.hnp && !orte_process_info.daemon &&
|
|
|
|
!orte_process_info.tool) {
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
}
|
|
|
|
|
2009-02-06 18:41:53 +03:00
|
|
|
/* 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 18:53:58 +04:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
2008-05-06 02:32:25 +04:00
|
|
|
"%s routed_binomial_update: %s --> %s",
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
ORTE_NAME_PRINT(target),
|
|
|
|
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 21:53:37 +04:00
|
|
|
/* if this is from a different job family, then I need to
|
|
|
|
* track how to send messages to it
|
2008-05-06 02:32:25 +04:00
|
|
|
*/
|
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 21:53:37 +04:00
|
|
|
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
|
2008-05-06 02:32:25 +04:00
|
|
|
*/
|
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 21:53:37 +04:00
|
|
|
if (orte_process_info.daemon) {
|
2008-05-06 02:32:25 +04:00
|
|
|
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 21:53:37 +04:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
|
|
|
"%s routed_binomial_update: diff job family routing job %s --> %s",
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
ORTE_JOBID_PRINT(target->jobid),
|
|
|
|
ORTE_NAME_PRINT(route)));
|
|
|
|
|
|
|
|
/* see if this target is already present - it will have a wildcard vpid,
|
|
|
|
* so we have to look for it with that condition
|
2008-05-06 02:32:25 +04:00
|
|
|
*/
|
2008-11-01 00:10:00 +03:00
|
|
|
rc = opal_hash_table_get_value_uint32(&jobfam_list,
|
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 21:53:37 +04:00
|
|
|
ORTE_JOB_FAMILY(target->jobid),
|
|
|
|
(void**)&route_copy);
|
|
|
|
if (ORTE_SUCCESS == rc && NULL != route_copy) {
|
2008-07-22 01:37:09 +04:00
|
|
|
/* target already present - update the route info
|
|
|
|
* in case it has changed
|
|
|
|
*/
|
|
|
|
*route_copy = *route;
|
2008-11-01 00:10:00 +03:00
|
|
|
rc = opal_hash_table_set_value_uint32(&jobfam_list,
|
2008-07-22 01:37:09 +04:00
|
|
|
ORTE_JOB_FAMILY(target->jobid), route_copy);
|
|
|
|
if (ORTE_SUCCESS != rc) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
}
|
|
|
|
return rc;
|
2008-05-06 02:32:25 +04:00
|
|
|
}
|
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 21:53:37 +04:00
|
|
|
|
|
|
|
/* not there, so add the route FOR THE JOB FAMILY*/
|
2008-09-08 19:39:30 +04:00
|
|
|
route_copy = (orte_process_name_t *) malloc(sizeof(orte_process_name_t));
|
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 21:53:37 +04:00
|
|
|
*route_copy = *route;
|
2008-11-01 00:10:00 +03:00
|
|
|
rc = opal_hash_table_set_value_uint32(&jobfam_list,
|
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 21:53:37 +04:00
|
|
|
ORTE_JOB_FAMILY(target->jobid), route_copy);
|
|
|
|
if (ORTE_SUCCESS != rc) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
}
|
|
|
|
return rc;
|
2008-05-06 02:32:25 +04:00
|
|
|
}
|
|
|
|
|
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 21:53:37 +04:00
|
|
|
/* THIS CAME FROM OUR OWN JOB FAMILY... */
|
|
|
|
|
2008-11-01 00:10:00 +03:00
|
|
|
opal_output(0, "%s CALL TO UPDATE ROUTE FOR OWN JOB FAMILY", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
|
2008-07-22 01:37:09 +04:00
|
|
|
|
2008-05-06 02:32:25 +04:00
|
|
|
return ORTE_ERR_NOT_SUPPORTED;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static orte_process_name_t get_route(orte_process_name_t *target)
|
|
|
|
{
|
2008-11-01 00:10:00 +03:00
|
|
|
orte_process_name_t *ret, daemon;
|
2008-05-06 02:32:25 +04:00
|
|
|
int rc;
|
|
|
|
|
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 21:53:37 +04:00
|
|
|
if (target->jobid == ORTE_JOBID_INVALID ||
|
|
|
|
target->vpid == ORTE_VPID_INVALID) {
|
|
|
|
ret = ORTE_NAME_INVALID;
|
|
|
|
goto found;
|
|
|
|
}
|
|
|
|
|
2008-05-06 02:32:25 +04:00
|
|
|
/* if it is me, then the route is just direct */
|
|
|
|
if (OPAL_EQUAL == opal_dss.compare(ORTE_PROC_MY_NAME, target, ORTE_NAME)) {
|
|
|
|
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 21:53:37 +04:00
|
|
|
/* if I am an application process, always route via my local daemon */
|
|
|
|
if (!orte_process_info.hnp && !orte_process_info.daemon &&
|
|
|
|
!orte_process_info.tool) {
|
|
|
|
ret = ORTE_PROC_MY_DAEMON;
|
|
|
|
goto found;
|
|
|
|
}
|
|
|
|
|
2008-11-01 00:10:00 +03:00
|
|
|
/****** HNP AND DAEMONS ONLY ******/
|
|
|
|
|
2009-02-06 18:41:53 +03:00
|
|
|
/* 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 21:53:37 +04:00
|
|
|
/* 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 */
|
|
|
|
if (orte_process_info.daemon) {
|
|
|
|
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
|
|
|
|
*/
|
2008-11-01 00:10:00 +03:00
|
|
|
rc = opal_hash_table_get_value_uint32(&jobfam_list,
|
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 21:53:37 +04:00
|
|
|
ORTE_JOB_FAMILY(target->jobid), (void**)&ret);
|
|
|
|
if (ORTE_SUCCESS == rc) {
|
|
|
|
/* got a good result - return it */
|
|
|
|
goto found;
|
|
|
|
}
|
|
|
|
/* not found - so we have no route */
|
|
|
|
ret = ORTE_NAME_INVALID;
|
|
|
|
goto found;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* THIS CAME FROM OUR OWN JOB FAMILY... */
|
|
|
|
|
2008-11-01 00:10:00 +03:00
|
|
|
/* if this is going to the HNP, send direct */
|
|
|
|
if (ORTE_PROC_MY_HNP->jobid == target->jobid &&
|
|
|
|
ORTE_PROC_MY_HNP->vpid == target->vpid) {
|
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_routed_base_output,
|
|
|
|
"%s routing not enabled - going direct",
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
ret = target;
|
2008-05-06 02:32:25 +04:00
|
|
|
goto found;
|
|
|
|
}
|
|
|
|
|
2008-11-01 00:10:00 +03:00
|
|
|
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 18:35:50 +03:00
|
|
|
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
|
|
|
|
ret = ORTE_NAME_INVALID;
|
2008-05-06 02:32:25 +04:00
|
|
|
goto found;
|
|
|
|
}
|
|
|
|
|
2008-11-01 00:10:00 +03:00
|
|
|
/* if the daemon is me, then send direct to the target! */
|
|
|
|
if (ORTE_PROC_MY_NAME->vpid == daemon.vpid) {
|
|
|
|
ret = target;
|
|
|
|
} else {
|
|
|
|
/* otherwise, we send it directly to that daemon */
|
|
|
|
ret = &daemon;
|
|
|
|
}
|
2008-05-06 02:32:25 +04:00
|
|
|
|
|
|
|
found:
|
2008-06-09 18:53:58 +04:00
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_routed_base_output,
|
2008-05-06 02:32:25 +04:00
|
|
|
"%s routed_binomial_get(%s) --> %s",
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
ORTE_NAME_PRINT(target),
|
|
|
|
ORTE_NAME_PRINT(ret)));
|
|
|
|
|
|
|
|
return *ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int process_callback(orte_jobid_t job, opal_buffer_t *buffer)
|
|
|
|
{
|
|
|
|
orte_proc_t **procs;
|
|
|
|
orte_job_t *jdata;
|
|
|
|
orte_std_cntr_t cnt;
|
|
|
|
char *rml_uri;
|
|
|
|
orte_process_name_t name;
|
|
|
|
int rc;
|
|
|
|
|
|
|
|
/* lookup the job object for this process */
|
|
|
|
if (NULL == (jdata = orte_get_job_data_object(job))) {
|
|
|
|
ORTE_ERROR_LOG(ORTE_ERR_NOT_FOUND);
|
|
|
|
return ORTE_ERR_NOT_FOUND;
|
|
|
|
}
|
|
|
|
procs = (orte_proc_t**)jdata->procs->addr;
|
|
|
|
|
|
|
|
/* unpack the data for each entry */
|
|
|
|
cnt = 1;
|
|
|
|
while (ORTE_SUCCESS == (rc = opal_dss.unpack(buffer, &rml_uri, &cnt, OPAL_STRING))) {
|
|
|
|
|
2008-06-09 18:53:58 +04:00
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_routed_base_output,
|
2008-05-06 02:32:25 +04:00
|
|
|
"%s routed_binomial:callback got uri %s",
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
(NULL == rml_uri) ? "NULL" : rml_uri));
|
|
|
|
|
|
|
|
if (rml_uri == NULL) continue;
|
|
|
|
|
|
|
|
/* we don't need to set the contact info into our rml
|
|
|
|
* hash table as we won't talk to the proc directly
|
|
|
|
*/
|
|
|
|
|
|
|
|
/* extract the proc's name */
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml_base_parse_uris(rml_uri, &name, NULL))) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
free(rml_uri);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
/* the procs are stored in vpid order, so update the record */
|
|
|
|
procs[name.vpid]->rml_uri = strdup(rml_uri);
|
|
|
|
free(rml_uri);
|
|
|
|
|
|
|
|
/* update the proc state */
|
|
|
|
if (procs[name.vpid]->state < ORTE_PROC_STATE_RUNNING) {
|
|
|
|
procs[name.vpid]->state = ORTE_PROC_STATE_RUNNING;
|
|
|
|
}
|
|
|
|
|
|
|
|
++jdata->num_reported;
|
|
|
|
cnt = 1;
|
|
|
|
}
|
|
|
|
if (ORTE_ERR_UNPACK_READ_PAST_END_OF_BUFFER != rc) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* if all procs have reported, update our job state */
|
|
|
|
if (jdata->num_reported == jdata->num_procs) {
|
|
|
|
/* update the job state */
|
|
|
|
if (jdata->state < ORTE_JOB_STATE_RUNNING) {
|
|
|
|
jdata->state = ORTE_JOB_STATE_RUNNING;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
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 21:53:37 +04:00
|
|
|
/* 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);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2008-05-06 02:32:25 +04:00
|
|
|
static int init_routes(orte_jobid_t job, opal_buffer_t *ndat)
|
|
|
|
{
|
|
|
|
/* the binomial module routes all proc communications through
|
|
|
|
* 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;
|
|
|
|
|
|
|
|
/* if I am a tool, then I stand alone - there is nothing to do */
|
|
|
|
if (orte_process_info.tool) {
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* 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
|
|
|
|
*/
|
|
|
|
if (orte_process_info.daemon) {
|
|
|
|
|
2008-06-09 18:53:58 +04:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
2008-05-06 02:32:25 +04:00
|
|
|
"%s routed_binomial: init routes for daemon job %s\n\thnp_uri %s",
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
ORTE_JOBID_PRINT(job),
|
|
|
|
(NULL == orte_process_info.my_hnp_uri) ? "NULL" : orte_process_info.my_hnp_uri));
|
|
|
|
|
|
|
|
if (NULL == ndat) {
|
|
|
|
/* indicates this is being called during orte_init.
|
|
|
|
* Get the HNP's name for possible later use
|
|
|
|
*/
|
|
|
|
if (NULL == orte_process_info.my_hnp_uri) {
|
|
|
|
/* fatal error */
|
|
|
|
ORTE_ERROR_LOG(ORTE_ERR_FATAL);
|
|
|
|
return ORTE_ERR_FATAL;
|
|
|
|
}
|
|
|
|
/* set the contact info into the hash table */
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml.set_contact_info(orte_process_info.my_hnp_uri))) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
return(rc);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* extract the hnp name and store it */
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml_base_parse_uris(orte_process_info.my_hnp_uri,
|
|
|
|
ORTE_PROC_MY_HNP, NULL))) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
2008-11-01 00:10:00 +03:00
|
|
|
/* set our lifeline to the HNP - we will abort if that connection is lost */
|
2008-05-06 02:32:25 +04:00
|
|
|
lifeline = ORTE_PROC_MY_HNP;
|
|
|
|
|
|
|
|
/* 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);
|
|
|
|
}
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
2008-06-09 18:53:58 +04:00
|
|
|
OPAL_OUTPUT_VERBOSE((2, orte_routed_base_output,
|
2008-05-06 02:32:25 +04:00
|
|
|
"%s routed_binomial: completed init routes",
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
if (orte_process_info.hnp) {
|
|
|
|
|
2008-06-09 18:53:58 +04:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
2008-05-06 02:32:25 +04:00
|
|
|
"%s routed_binomial: init routes for HNP job %s",
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
ORTE_JOBID_PRINT(job)));
|
|
|
|
|
|
|
|
if (NULL == ndat) {
|
|
|
|
/* if ndat is NULL, then this is being called during init, so just
|
|
|
|
* make myself available to catch any reported contact info
|
|
|
|
*/
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_routed_base_comm_start())) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
/* the HNP has no lifeline */
|
|
|
|
lifeline = NULL;
|
|
|
|
} else {
|
|
|
|
/* if this is for my own jobid, then I am getting an update of RML info
|
|
|
|
* for the daemons - so update our contact info and routes
|
|
|
|
*/
|
|
|
|
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 */
|
|
|
|
if (ORTE_SUCCESS != (rc = process_callback(job, ndat))) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
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) {
|
|
|
|
int rc;
|
|
|
|
|
2008-06-09 18:53:58 +04:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
2008-05-06 02:32:25 +04:00
|
|
|
"%s routed_binomial: init routes w/non-NULL data",
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
2009-02-06 18:41:53 +03:00
|
|
|
/* 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 18:35:50 +03:00
|
|
|
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_binomial_init_routes: diff job family - sending update to %s",
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_HNP)));
|
|
|
|
|
|
|
|
if (0 > (rc = orte_rml.send_buffer(ORTE_PROC_MY_HNP, ndat,
|
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 21:53:37 +04:00
|
|
|
ORTE_RML_TAG_RML_INFO_UPDATE, 0))) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
return rc;
|
|
|
|
}
|
2008-11-18 18:35:50 +03:00
|
|
|
|
|
|
|
/* 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 21:53:37 +04:00
|
|
|
* 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_binomial_init_routes: ack recvd",
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME)));
|
|
|
|
|
2008-11-18 18:35:50 +03:00
|
|
|
/* 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 21:53:37 +04:00
|
|
|
*/
|
|
|
|
}
|
2008-05-06 02:32:25 +04:00
|
|
|
return ORTE_SUCCESS;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* 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 18:53:58 +04:00
|
|
|
OPAL_OUTPUT_VERBOSE((1, orte_routed_base_output,
|
2008-05-06 02:32:25 +04:00
|
|
|
"%s routed_binomial: init routes for proc job %s\n\thnp_uri %s\n\tdaemon uri %s",
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), ORTE_JOBID_PRINT(job),
|
|
|
|
(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));
|
|
|
|
|
|
|
|
if (NULL == orte_process_info.my_daemon_uri) {
|
|
|
|
/* in this module, we absolutely MUST have this information - if
|
|
|
|
* we didn't get it, then error out
|
|
|
|
*/
|
2008-06-09 18:53:58 +04:00
|
|
|
opal_output(0, "%s ERROR: Failed to identify the local daemon's URI",
|
2008-05-06 02:32:25 +04:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
|
2008-06-09 18:53:58 +04:00
|
|
|
opal_output(0, "%s ERROR: This is a fatal condition when the binomial router",
|
2008-05-06 02:32:25 +04:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
|
2008-06-09 18:53:58 +04:00
|
|
|
opal_output(0, "%s ERROR: has been selected - either select the unity router",
|
2008-05-06 02:32:25 +04:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
|
2008-06-09 18:53:58 +04:00
|
|
|
opal_output(0, "%s ERROR: or ensure that the local daemon info is provided",
|
2008-05-06 02:32:25 +04:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME));
|
|
|
|
return ORTE_ERR_FATAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* 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
|
|
|
|
*/
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml_base_parse_uris(orte_process_info.my_hnp_uri,
|
|
|
|
ORTE_PROC_MY_HNP, NULL))) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* 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
|
|
|
|
*/
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml.set_contact_info(orte_process_info.my_daemon_uri))) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
return(rc);
|
|
|
|
}
|
|
|
|
/* extract the daemon's name so we can update the routing table */
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml_base_parse_uris(orte_process_info.my_daemon_uri,
|
|
|
|
ORTE_PROC_MY_DAEMON, NULL))) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* set our lifeline to the local daemon - we will abort if this connection is lost */
|
|
|
|
lifeline = ORTE_PROC_MY_DAEMON;
|
|
|
|
|
|
|
|
/* 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.
|
|
|
|
*/
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_routed_base_register_sync(true))) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
/* no answer is expected or coming */
|
|
|
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
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 18:53:58 +04:00
|
|
|
opal_output(0, "%s routed:binomial: Connection to lifeline %s lost",
|
2008-05-06 02:32:25 +04:00
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
|
|
|
ORTE_NAME_PRINT(lifeline));
|
|
|
|
return ORTE_ERR_FATAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* we don't care about this one, so return success */
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static bool route_is_defined(const orte_process_name_t *target)
|
|
|
|
{
|
2008-11-01 00:10:00 +03:00
|
|
|
/* find out what daemon hosts this proc */
|
|
|
|
if (ORTE_VPID_INVALID == orte_ess.proc_get_daemon((orte_process_name_t*)target)) {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2008-05-06 02:32:25 +04:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2008-12-21 02:23:11 +03:00
|
|
|
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;
|
|
|
|
}
|
|
|
|
|
2008-10-03 00:08:27 +04:00
|
|
|
static int binomial_tree(int rank, int parent, int me, int num_procs,
|
|
|
|
int *nchildren, opal_list_t *childrn, opal_bitmap_t *relatives)
|
2008-05-06 02:32:25 +04:00
|
|
|
{
|
|
|
|
int i, bitmap, peer, hibit, mask, found;
|
2008-10-03 00:08:27 +04:00
|
|
|
orte_routed_tree_t *child;
|
|
|
|
opal_bitmap_t *relations;
|
2008-05-06 02:32:25 +04:00
|
|
|
|
|
|
|
/* is this me? */
|
|
|
|
if (me == rank) {
|
|
|
|
bitmap = opal_cube_dim(num_procs);
|
|
|
|
|
|
|
|
hibit = opal_hibit(rank, bitmap);
|
|
|
|
--bitmap;
|
|
|
|
|
|
|
|
for (i = hibit + 1, mask = 1 << i; i <= bitmap; ++i, mask <<= 1) {
|
|
|
|
peer = rank | mask;
|
|
|
|
if (peer < num_procs) {
|
2008-10-03 00:08:27 +04:00
|
|
|
child = OBJ_NEW(orte_routed_tree_t);
|
|
|
|
child->vpid = peer;
|
2008-06-09 18:53:58 +04:00
|
|
|
OPAL_OUTPUT_VERBOSE((3, orte_routed_base_output,
|
2008-05-06 02:32:25 +04:00
|
|
|
"%s routed:binomial found child %s",
|
|
|
|
ORTE_NAME_PRINT(ORTE_PROC_MY_NAME),
|
2008-10-03 00:08:27 +04:00
|
|
|
ORTE_VPID_PRINT(child->vpid)));
|
|
|
|
if (NULL != childrn) {
|
|
|
|
/* this is a direct child - add it to my list */
|
|
|
|
opal_list_append(childrn, &child->super);
|
|
|
|
(*nchildren)++;
|
|
|
|
/* setup the relatives bitmap */
|
|
|
|
opal_bitmap_init(&child->relatives, num_procs);
|
|
|
|
/* point to the relatives */
|
|
|
|
relations = &child->relatives;
|
|
|
|
} else {
|
|
|
|
/* we are recording someone's relatives - set the bit */
|
|
|
|
opal_bitmap_set_bit(relatives, peer);
|
|
|
|
/* point to this relations */
|
|
|
|
relations = relatives;
|
|
|
|
}
|
|
|
|
/* search for this child's relatives */
|
|
|
|
binomial_tree(0, 0, peer, num_procs, NULL, NULL, relations);
|
2008-05-06 02:32:25 +04:00
|
|
|
}
|
|
|
|
}
|
|
|
|
return parent;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* find the children of this rank */
|
|
|
|
bitmap = opal_cube_dim(num_procs);
|
|
|
|
|
|
|
|
hibit = opal_hibit(rank, bitmap);
|
|
|
|
--bitmap;
|
|
|
|
|
|
|
|
for (i = hibit + 1, mask = 1 << i; i <= bitmap; ++i, mask <<= 1) {
|
|
|
|
peer = rank | mask;
|
|
|
|
if (peer < num_procs) {
|
|
|
|
/* execute compute on this child */
|
2008-10-03 00:08:27 +04:00
|
|
|
if (0 <= (found = binomial_tree(peer, rank, me, num_procs, nchildren, childrn, relatives))) {
|
2008-05-06 02:32:25 +04:00
|
|
|
return found;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return -1;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int update_routing_tree(void)
|
|
|
|
{
|
2008-10-03 17:41:08 +04:00
|
|
|
orte_routed_tree_t *child;
|
|
|
|
int j;
|
2008-05-06 02:32:25 +04:00
|
|
|
opal_list_item_t *item;
|
|
|
|
|
2008-06-20 07:26:13 +04:00
|
|
|
/* if I am anything other than a daemon or the HNP, this
|
|
|
|
* is a meaningless command as I am not allowed to route
|
|
|
|
*/
|
|
|
|
if (!orte_process_info.daemon && !orte_process_info.hnp) {
|
|
|
|
return ORTE_ERR_NOT_SUPPORTED;
|
|
|
|
}
|
|
|
|
|
2008-05-06 02:32:25 +04:00
|
|
|
/* clear the list of children if any are already present */
|
|
|
|
while (NULL != (item = opal_list_remove_first(&my_children))) {
|
|
|
|
OBJ_RELEASE(item);
|
|
|
|
}
|
|
|
|
num_children = 0;
|
|
|
|
|
2008-10-03 00:08:27 +04:00
|
|
|
/* compute my direct children and the bitmap that shows which vpids
|
|
|
|
* lie underneath their branch
|
|
|
|
*/
|
2008-05-06 02:32:25 +04:00
|
|
|
my_parent.vpid = binomial_tree(0, 0, ORTE_PROC_MY_NAME->vpid,
|
2008-10-03 00:08:27 +04:00
|
|
|
orte_process_info.num_procs,
|
|
|
|
&num_children, &my_children, NULL);
|
2008-05-06 02:32:25 +04:00
|
|
|
|
2008-10-03 17:41:08 +04:00
|
|
|
if (0 < opal_output_get_verbosity(orte_routed_base_output)) {
|
2008-10-03 00:08:27 +04:00
|
|
|
opal_output(0, "%s: parent %d num_children %d", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), my_parent.vpid, num_children);
|
|
|
|
for (item = opal_list_get_first(&my_children);
|
|
|
|
item != opal_list_get_end(&my_children);
|
|
|
|
item = opal_list_get_next(item)) {
|
|
|
|
child = (orte_routed_tree_t*)item;
|
|
|
|
opal_output(0, "%s: \tchild %d", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), child->vpid);
|
|
|
|
for (j=0; j < (int)orte_process_info.num_procs; j++) {
|
|
|
|
if (opal_bitmap_is_set_bit(&child->relatives, j)) {
|
|
|
|
opal_output(0, "%s: \t\trelation %d", ORTE_NAME_PRINT(ORTE_PROC_MY_NAME), j);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-05-06 02:32:25 +04:00
|
|
|
return ORTE_SUCCESS;
|
|
|
|
}
|
|
|
|
|
2009-01-27 22:13:56 +03:00
|
|
|
static orte_vpid_t get_routing_tree(opal_list_t *children)
|
2008-05-06 02:32:25 +04:00
|
|
|
{
|
|
|
|
opal_list_item_t *item;
|
2009-01-27 22:13:56 +03:00
|
|
|
orte_routed_tree_t *child, *nm;
|
2008-05-06 02:32:25 +04:00
|
|
|
|
2008-06-20 07:26:13 +04:00
|
|
|
/* if I am anything other than a daemon or the HNP, this
|
|
|
|
* is a meaningless command as I am not allowed to route
|
|
|
|
*/
|
|
|
|
if (!orte_process_info.daemon && !orte_process_info.hnp) {
|
|
|
|
return ORTE_VPID_INVALID;
|
|
|
|
}
|
|
|
|
|
2008-05-06 02:32:25 +04:00
|
|
|
/* the binomial routing tree always goes to our children,
|
|
|
|
* for any job
|
|
|
|
*/
|
2008-05-07 00:16:17 +04:00
|
|
|
if (NULL != children) {
|
|
|
|
for (item = opal_list_get_first(&my_children);
|
|
|
|
item != opal_list_get_end(&my_children);
|
|
|
|
item = opal_list_get_next(item)) {
|
2008-10-03 00:08:27 +04:00
|
|
|
child = (orte_routed_tree_t*)item;
|
2009-01-27 22:13:56 +03:00
|
|
|
nm = OBJ_NEW(orte_routed_tree_t);
|
|
|
|
nm->vpid = child->vpid;
|
|
|
|
opal_bitmap_copy(&nm->relatives, &child->relatives);
|
|
|
|
opal_list_append(children, &nm->super);
|
2008-05-07 00:16:17 +04:00
|
|
|
}
|
2008-05-06 02:32:25 +04:00
|
|
|
}
|
2008-11-01 00:10:00 +03:00
|
|
|
|
2008-05-06 02:32:25 +04:00
|
|
|
/* return my parent's vpid */
|
|
|
|
return my_parent.vpid;
|
|
|
|
}
|
|
|
|
|
2008-11-01 00:10:00 +03:00
|
|
|
static int get_wireup_info(opal_buffer_t *buf)
|
|
|
|
{
|
|
|
|
int rc;
|
|
|
|
|
|
|
|
/* if I am anything other than the HNP, this
|
|
|
|
* is a meaningless command as I cannot get
|
|
|
|
* the requested info
|
|
|
|
*/
|
|
|
|
if (!orte_process_info.hnp) {
|
|
|
|
return ORTE_ERR_NOT_SUPPORTED;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* 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;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (ORTE_SUCCESS != (rc = orte_rml_base_get_contact_info(ORTE_PROC_MY_NAME->jobid, buf))) {
|
|
|
|
ORTE_ERROR_LOG(rc);
|
|
|
|
OBJ_RELEASE(buf);
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
|
|
|
return ORTE_SUCCESS;
|
|
|
|
}
|
|
|
|
|
2008-10-03 00:08:27 +04:00
|
|
|
|
2008-05-06 02:32:25 +04:00
|
|
|
#if OPAL_ENABLE_FT == 1
|
|
|
|
static int binomial_ft_event(int state)
|
|
|
|
{
|
|
|
|
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
|
|
|
|
|