1
1
openmpi/opal/mca/pmix/native/pmix_native.c

1358 строки
46 KiB
C
Исходник Обычный вид История

Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2014 Intel, Inc. All rights reserved.
* Copyright (c) 2014 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* Copyright (c) 2014 Mellanox Technologies, Inc.
* All rights reserved.
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "opal_config.h"
#include "opal/constants.h"
#include "opal/types.h"
#ifdef HAVE_STRING_H
#include <string.h>
#endif
#ifdef HAVE_UNISTD_H
#include <unistd.h>
#endif
#include "opal/dss/dss.h"
#include "opal/mca/event/event.h"
#include "opal/mca/hwloc/base/base.h"
#include "opal/runtime/opal.h"
#include "opal/runtime/opal_progress_threads.h"
#include "opal/util/error.h"
#include "opal/util/output.h"
#include "opal/util/proc.h"
#include "opal/util/show_help.h"
#include "opal/mca/pmix/base/base.h"
#include "pmix_native.h"
#include "opal/mca/dstore/base/base.h"
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
static int native_init(void);
static int native_fini(void);
static bool native_initialized(void);
static int native_abort(int flag, const char msg[]);
static int native_fence(opal_process_name_t *procs, size_t nprocs);
static int native_fence_nb(opal_process_name_t *procs, size_t nprocs,
opal_pmix_cbfunc_t cbfunc, void *cbdata);
static int native_put(opal_pmix_scope_t scope,
opal_value_t *kv);
static int native_get(const opal_process_name_t *id,
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
const char *key,
opal_value_t **kv);
static void native_get_nb(const opal_process_name_t *id,
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
const char *key,
opal_pmix_cbfunc_t cbfunc,
void *cbdata);
static int native_publish(const char service_name[],
opal_list_t *info,
const char port[]);
static int native_lookup(const char service_name[],
opal_list_t *info,
char port[], int portLen);
static int native_unpublish(const char service_name[],
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
opal_list_t *info);
static bool native_get_attr(const char *attr, opal_value_t **kv);
static int native_get_attr_nb(const char *attr,
opal_pmix_cbfunc_t cbfunc,
void *cbdata);
static int native_spawn(int count, const char * cmds[],
int argcs[], const char ** argvs[],
const int maxprocs[],
opal_list_t *info_keyval_vector,
opal_list_t *preput_keyval_vector,
char jobId[], int jobIdSize,
int errors[]);
static int native_job_connect(const char jobId[]);
static int native_job_disconnect(const char jobId[]);
const opal_pmix_base_module_t opal_pmix_native_module = {
native_init,
native_fini,
native_initialized,
native_abort,
native_fence,
native_fence_nb,
native_put,
native_get,
native_get_nb,
native_publish,
native_lookup,
native_unpublish,
native_get_attr,
native_get_attr_nb,
native_spawn,
native_job_connect,
native_job_disconnect,
opal_pmix_base_register_handler,
opal_pmix_base_deregister_handler
};
// local variables
static int init_cntr = 0;
opal_process_name_t native_pname;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
static char *local_uri = NULL;
static uint32_t sm_flag;
static void unpack_segment_info(opal_buffer_t *buf, opal_process_name_t *id, char** seg_info)
{
int cnt;
int rc;
char *sinfo;
opal_process_name_t uid;
*seg_info = NULL;
/* extract the id of the contributor from the blob */
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(buf, &uid, &cnt, OPAL_NAME))) {
if (OPAL_ERR_UNPACK_READ_PAST_END_OF_BUFFER == rc) {
return;
}
OPAL_ERROR_LOG(rc);
return;
}
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(buf, &sinfo, &cnt, OPAL_STRING))) {
OPAL_ERROR_LOG(rc);
return;
}
*id = uid;
*seg_info = sinfo;
}
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
/* callback for wait completion */
static void wait_cbfunc(opal_buffer_t *buf, void *cbdata)
{
pmix_cb_t *cb = (pmix_cb_t*)cbdata;
int status=OPAL_SUCCESS;
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native recv callback activated with %d bytes",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
(NULL == buf) ? -1 : (int)buf->bytes_used);
if (NULL != buf) {
/* transfer the data to the cb */
opal_dss.copy_payload(&cb->data, buf);
}
if (NULL != cb->cbfunc) {
cb->cbfunc(status, NULL, cb->cbdata);
}
cb->active = false;
}
static int pmix_sm_attach(uint32_t jid, char *seg_info)
{
int rc;
opal_dstore_attr_t *attr;
opal_list_t attrs;
OBJ_CONSTRUCT(&attrs, opal_list_t);
attr = OBJ_NEW(opal_dstore_attr_t);
attr->jobid = jid;
attr->connection_info = strdup(seg_info);
opal_list_append(&attrs, &attr->super);
rc = opal_dstore.update(opal_dstore_modex, &attrs);
opal_list_remove_item(&attrs, &attr->super);
OBJ_RELEASE(attr);
OPAL_LIST_DESTRUCT(&attrs);
return rc;
}
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
static int native_init(void)
{
char **uri, *srv;
++init_cntr;
if (1 < init_cntr) {
return OPAL_SUCCESS;
}
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native init called",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
/* if we don't have a path to the daemon rendezvous point,
* then we need to return an error UNLESS we have been directed
* to allow init prior to having an identified server. This is
* needed for singletons as they will start without a server
* to support them, but may have one assigned at a later time */
if (NULL == mca_pmix_native_component.uri) {
if (!opal_pmix_base_allow_delayed_server) {
/* not ready yet, so decrement our init_cntr so we can come thru
* here again */
--init_cntr;
/* let the caller know that the server isn't available yet */
return OPAL_ERR_SERVER_NOT_AVAIL;
}
if (NULL == (srv = getenv("PMIX_SERVER_URI"))) {
/* error out - should have been here, but isn't */
return OPAL_ERROR;
}
mca_pmix_native_component.uri = strdup(srv);
mca_pmix_native_component.id = OPAL_PROC_MY_NAME;
}
/* if we have it, setup the path to the daemon rendezvous point */
if (NULL != mca_pmix_native_component.uri) {
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native constructing component fields with server %s",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
mca_pmix_native_component.uri);
memset(&mca_pmix_native_component.address, 0, sizeof(struct sockaddr_un));
mca_pmix_native_component.address.sun_family = AF_UNIX;
uri = opal_argv_split(mca_pmix_native_component.uri, ':');
if (2 != opal_argv_count(uri)) {
return OPAL_ERROR;
}
/* if the rendezvous file doesn't exist, that's an error */
if (0 != access(uri[1], R_OK)) {
return OPAL_ERR_NOT_FOUND;
}
opal_convert_string_to_process_name(&mca_pmix_native_component.server, uri[0]);
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
snprintf(mca_pmix_native_component.address.sun_path,
sizeof(mca_pmix_native_component.address.sun_path)-1,
"%s", uri[1]);
opal_argv_free(uri);
/* create an event base and progress thread for us */
if (NULL == (mca_pmix_native_component.evbase = opal_start_progress_thread("pmix_native", true))) {
return OPAL_ERROR;
}
}
char* seg_info;
void *hdl;
int rc;
/* check if shared memory region is supported */
opal_dstore.get_handle(opal_dstore_modex, &hdl);
if(0 == strcmp("sm", ((opal_dstore_handle_t *)hdl)->storage_component->base_version.mca_component_name)) {
sm_flag = 1;
} else {
sm_flag = 0;
}
/* if shared memory segment is available, then attach to shared memory region created by pmix server */
if (1 == sm_flag) {
if (NULL == (seg_info = getenv("PMIX_SEG_INFO"))) {
/* error out - should have been here, but isn't */
return OPAL_ERROR;
}
rc = pmix_sm_attach(OPAL_PROC_MY_NAME.jobid, seg_info);
if (OPAL_SUCCESS != rc) {
/* error out - should have shared memory segment attached */
return OPAL_ERROR;
}
}
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
/* we will connect on first send */
return OPAL_SUCCESS;
}
static int native_fini(void)
{
opal_buffer_t *msg;
pmix_cb_t *cb;
pmix_cmd_t cmd = PMIX_FINALIZE_CMD;
int rc;
if (1 != init_cntr) {
--init_cntr;
return OPAL_SUCCESS;
}
init_cntr = 0;
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native finalize called",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
if (NULL == mca_pmix_native_component.uri) {
/* nothing was setup, so return */
return OPAL_SUCCESS;
}
if (PMIX_USOCK_CONNECTED == mca_pmix_native_component.state) {
/* setup a cmd message to notify the PMIx
* server that we are normally terminating */
msg = OBJ_NEW(opal_buffer_t);
/* pack the cmd */
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &cmd, 1, PMIX_CMD_T))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
/* create a callback object as we need to pass it to the
* recv routine so we know which callback to use when
* the return message is recvd */
cb = OBJ_NEW(pmix_cb_t);
cb->active = true;
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native sending finalize sync to server",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
/* push the message into our event base to send to the server */
PMIX_ACTIVATE_SEND_RECV(msg, wait_cbfunc, cb);
/* wait for the ack to return */
PMIX_WAIT_FOR_COMPLETION(cb->active);
OBJ_RELEASE(cb);
}
if (NULL != mca_pmix_native_component.evbase) {
opal_stop_progress_thread("pmix_native", true);
mca_pmix_native_component.evbase = NULL;
}
if (0 <= mca_pmix_native_component.sd) {
CLOSE_THE_SOCKET(mca_pmix_native_component.sd);
}
return OPAL_SUCCESS;
}
static bool native_initialized(void)
{
if (0 < init_cntr) {
return true;
}
return false;
}
static int native_abort(int flag, const char msg[])
{
opal_buffer_t *bfr;
pmix_cmd_t cmd = PMIX_ABORT_CMD;
int rc;
pmix_cb_t *cb;
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native abort called",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
if (NULL == mca_pmix_native_component.uri) {
/* no server available, so just return */
return OPAL_SUCCESS;
}
/* create a buffer to hold the message */
bfr = OBJ_NEW(opal_buffer_t);
/* pack the cmd */
if (OPAL_SUCCESS != (rc = opal_dss.pack(bfr, &cmd, 1, PMIX_CMD_T))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(bfr);
return rc;
}
/* pack the status flag */
if (OPAL_SUCCESS != (rc = opal_dss.pack(bfr, &flag, 1, OPAL_INT))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(bfr);
return rc;
}
/* pack the string message - a NULL is okay */
if (OPAL_SUCCESS != (rc = opal_dss.pack(bfr, &msg, 1, OPAL_STRING))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(bfr);
return rc;
}
/* create a callback object as we need to pass it to the
* recv routine so we know which callback to use when
* the return message is recvd */
cb = OBJ_NEW(pmix_cb_t);
cb->active = true;
/* push the message into our event base to send to the server */
PMIX_ACTIVATE_SEND_RECV(bfr, wait_cbfunc, cb);
/* wait for the release */
PMIX_WAIT_FOR_COMPLETION(cb->active);
OBJ_RELEASE(cb);
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
return OPAL_SUCCESS;
}
static int native_spawn(int count, const char * cmds[],
int argcs[], const char ** argvs[],
const int maxprocs[],
opal_list_t *info_keyval_vector,
opal_list_t *preput_keyval_vector,
char jobId[], int jobIdSize,
int errors[])
{
return OPAL_ERR_NOT_SUPPORTED;
}
static int native_put(opal_pmix_scope_t scope,
opal_value_t *kv)
{
int rc;
/* pack the cache that matches the scope */
if (PMIX_LOCAL == scope) {
if (NULL == mca_pmix_native_component.cache_local) {
mca_pmix_native_component.cache_local = OBJ_NEW(opal_buffer_t);
}
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native put local data for key %s",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME), kv->key);
if (OPAL_SUCCESS != (rc = opal_dss.pack(mca_pmix_native_component.cache_local, &kv, 1, OPAL_VALUE))) {
OPAL_ERROR_LOG(rc);
}
} else if (PMIX_REMOTE == scope) {
if (NULL == mca_pmix_native_component.cache_remote) {
mca_pmix_native_component.cache_remote = OBJ_NEW(opal_buffer_t);
}
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native put remote data for key %s",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME), kv->key);
if (OPAL_SUCCESS != (rc = opal_dss.pack(mca_pmix_native_component.cache_remote, &kv, 1, OPAL_VALUE))) {
OPAL_ERROR_LOG(rc);
}
} else {
/* must be global */
if (NULL == mca_pmix_native_component.cache_global) {
mca_pmix_native_component.cache_global = OBJ_NEW(opal_buffer_t);
}
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native put global data for key %s",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME), kv->key);
if (OPAL_SUCCESS != (rc = opal_dss.pack(mca_pmix_native_component.cache_global, &kv, 1, OPAL_VALUE))) {
OPAL_ERROR_LOG(rc);
}
}
/* if this is our uri, save it as we need to send it to our server
* as a special, separate item */
if (0 == strcmp(OPAL_DSTORE_URI, kv->key)) {
local_uri = strdup(kv->data.string);
}
/* have to save a copy locally as some of our components will
* look for it */
(void)opal_dstore.store(opal_dstore_internal, &OPAL_PROC_MY_NAME, kv);
return rc;
}
static int native_fence(opal_process_name_t *procs, size_t nprocs)
{
opal_buffer_t *msg, *bptr;
pmix_cmd_t cmd = PMIX_FENCE_CMD;
pmix_cb_t *cb;
int rc, ret;
opal_pmix_scope_t scope;
int32_t cnt;
opal_value_t *kp;
opal_process_name_t id;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
size_t i;
uint64_t np;
char *seg_info;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native executing fence on %u procs",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME), (unsigned int)nprocs);
if (NULL == mca_pmix_native_component.uri) {
/* no server available, so just return */
return OPAL_SUCCESS;
}
msg = OBJ_NEW(opal_buffer_t);
/* pack the fence cmd */
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &cmd, 1, PMIX_CMD_T))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
/* pack the number of procs */
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &nprocs, 1, OPAL_SIZE))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
if (0 < nprocs) {
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, procs, nprocs, OPAL_NAME))) {
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
}
/* provide our URI */
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &local_uri, 1, OPAL_STRING))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
/* only do it once */
if (NULL != local_uri) {
free(local_uri);
local_uri = NULL;
}
/* pack 1 if we have sm dstore enabled, 0 otherwise */
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &sm_flag, 1, OPAL_UINT32))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
/* if we haven't already done it, ensure we have committed our values */
if (NULL != mca_pmix_native_component.cache_local) {
scope = PMIX_LOCAL;
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &scope, 1, PMIX_SCOPE_T))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &mca_pmix_native_component.cache_local, 1, OPAL_BUFFER))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
OBJ_RELEASE(mca_pmix_native_component.cache_local);
}
if (NULL != mca_pmix_native_component.cache_remote) {
scope = PMIX_REMOTE;
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &scope, 1, PMIX_SCOPE_T))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &mca_pmix_native_component.cache_remote, 1, OPAL_BUFFER))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
OBJ_RELEASE(mca_pmix_native_component.cache_remote);
}
if (NULL != mca_pmix_native_component.cache_global) {
scope = PMIX_GLOBAL;
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &scope, 1, PMIX_SCOPE_T))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &mca_pmix_native_component.cache_global, 1, OPAL_BUFFER))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
OBJ_RELEASE(mca_pmix_native_component.cache_global);
}
/* create a callback object as we need to pass it to the
* recv routine so we know which callback to use when
* the return message is recvd */
cb = OBJ_NEW(pmix_cb_t);
cb->active = true;
/* push the message into our event base to send to the server */
PMIX_ACTIVATE_SEND_RECV(msg, wait_cbfunc, cb);
/* wait for the fence to complete */
PMIX_WAIT_FOR_COMPLETION(cb->active);
/* get the number of contributors */
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(&cb->data, &np, &cnt, OPAL_UINT64))) {
OPAL_ERROR_LOG(rc);
return rc;
}
/* if data was returned, unpack and store it */
for (i=0; i < np; i++) {
if (0 == sm_flag) {
/* get the buffer that contains the data for the next proc */
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(&cb->data, &msg, &cnt, OPAL_BUFFER))) {
if (OPAL_ERR_UNPACK_READ_PAST_END_OF_BUFFER == rc) {
break;
}
OPAL_ERROR_LOG(rc);
return rc;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
}
/* extract the id of the contributor from the blob */
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(msg, &id, &cnt, OPAL_NAME))) {
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
OPAL_ERROR_LOG(rc);
return rc;
}
/* extract all blobs from this proc, starting with the scope */
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
cnt = 1;
while (OPAL_SUCCESS == (rc = opal_dss.unpack(msg, &scope, &cnt, PMIX_SCOPE_T))) {
/* extract the blob for this scope */
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(msg, &bptr, &cnt, OPAL_BUFFER))) {
OPAL_ERROR_LOG(rc);
return rc;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
}
/* now unpack and store the values - everything goes into our internal store */
cnt = 1;
while (OPAL_SUCCESS == (rc = opal_dss.unpack(bptr, &kp, &cnt, OPAL_VALUE))) {
if (OPAL_SUCCESS != (ret = opal_dstore.store(opal_dstore_internal, &id, kp))) {
OPAL_ERROR_LOG(ret);
}
OBJ_RELEASE(kp);
cnt = 1;
}
OBJ_RELEASE(bptr);
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
cnt = 1;
}
if (OPAL_ERR_UNPACK_READ_PAST_END_OF_BUFFER != rc) {
OPAL_ERROR_LOG(rc);
}
OBJ_RELEASE(msg);
} else {
unpack_segment_info(&cb->data, &id, &seg_info);
if (NULL != seg_info) {
pmix_sm_attach(id.jobid, seg_info);
}
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
}
if (OPAL_SUCCESS != rc && OPAL_ERR_UNPACK_READ_PAST_END_OF_BUFFER != rc) {
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
OPAL_ERROR_LOG(rc);
} else {
rc = OPAL_SUCCESS;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
}
}
OBJ_RELEASE(cb);
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native fence released",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
return OPAL_SUCCESS;
}
static void fencenb_cbfunc(opal_buffer_t *buf, void *cbdata)
{
pmix_cb_t *cb = (pmix_cb_t*)cbdata;
opal_buffer_t *msg, *bptr;
int rc, ret;
opal_pmix_scope_t scope;
int32_t cnt;
opal_value_t *kp;
opal_process_name_t id;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
size_t i;
uint64_t np;
char *seg_info;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
/* get the number of contributors */
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(buf, &np, &cnt, OPAL_UINT64))) {
OPAL_ERROR_LOG(rc);
return;
}
/* if data was returned, unpack and store it */
for (i=0; i < np; i++) {
if (0 == sm_flag) {
/* get the buffer that contains the data for the next proc */
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(buf, &msg, &cnt, OPAL_BUFFER))) {
if (OPAL_ERR_UNPACK_READ_PAST_END_OF_BUFFER == rc) {
break;
}
OPAL_ERROR_LOG(rc);
return;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
}
/* extract the id of the contributor from the blob */
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(msg, &id, &cnt, OPAL_NAME))) {
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
OPAL_ERROR_LOG(rc);
return;
}
/* extract all blobs from this proc, starting with the scope */
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
cnt = 1;
while (OPAL_SUCCESS == (rc = opal_dss.unpack(msg, &scope, &cnt, PMIX_SCOPE_T))) {
/* extract the blob for this scope */
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(msg, &bptr, &cnt, OPAL_BUFFER))) {
OPAL_ERROR_LOG(rc);
return;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
}
/* now unpack and store the values - everything goes into our internal store */
cnt = 1;
while (OPAL_SUCCESS == (rc = opal_dss.unpack(bptr, &kp, &cnt, OPAL_VALUE))) {
if (OPAL_SUCCESS != (ret = opal_dstore.store(opal_dstore_internal, &id, kp))) {
OPAL_ERROR_LOG(ret);
}
OBJ_RELEASE(kp);
cnt = 1;
}
OBJ_RELEASE(bptr);
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
cnt = 1;
}
if (OPAL_ERR_UNPACK_READ_PAST_END_OF_BUFFER != rc) {
OPAL_ERROR_LOG(rc);
}
OBJ_RELEASE(msg);
} else {
unpack_segment_info(buf, &id, &seg_info);
if (NULL != seg_info) {
pmix_sm_attach(id.jobid, seg_info);
}
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
}
if (OPAL_SUCCESS != rc && OPAL_ERR_UNPACK_READ_PAST_END_OF_BUFFER != rc) {
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
OPAL_ERROR_LOG(rc);
}
}
/* if a callback was provided, execute it */
if (NULL != cb && NULL != cb->cbfunc) {
cb->cbfunc(rc, NULL, cb->cbdata);
}
OBJ_RELEASE(cb);
}
static int native_fence_nb(opal_process_name_t *procs, size_t nprocs,
opal_pmix_cbfunc_t cbfunc, void *cbdata)
{
opal_buffer_t *msg;
pmix_cmd_t cmd = PMIX_FENCENB_CMD;
int rc;
pmix_cb_t *cb;
opal_pmix_scope_t scope;
if (NULL == mca_pmix_native_component.uri) {
/* no server available, so just execute the callback */
if (NULL != cbfunc) {
cbfunc(OPAL_SUCCESS, NULL, cbdata);
}
return OPAL_SUCCESS;
}
msg = OBJ_NEW(opal_buffer_t);
/* pack the fence cmd */
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &cmd, 1, PMIX_CMD_T))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
/* pack the number of procs */
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &nprocs, 1, OPAL_SIZE))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
if (0 < nprocs) {
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, procs, nprocs, OPAL_NAME))) {
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
}
/* provide our URI */
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &local_uri, 1, OPAL_STRING))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
/* only do it once */
if (NULL != local_uri) {
free(local_uri);
local_uri = NULL;
}
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
/* pack 1 if we have sm dstore enabled, 0 otherwise */
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &sm_flag, 1, OPAL_UINT32))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
/* if we haven't already done it, ensure we have committed our values */
if (NULL != mca_pmix_native_component.cache_local) {
scope = PMIX_LOCAL;
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &scope, 1, PMIX_SCOPE_T))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &mca_pmix_native_component.cache_local, 1, OPAL_BUFFER))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
OBJ_RELEASE(mca_pmix_native_component.cache_local);
}
if (NULL != mca_pmix_native_component.cache_remote) {
scope = PMIX_REMOTE;
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &scope, 1, PMIX_SCOPE_T))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &mca_pmix_native_component.cache_remote, 1, OPAL_BUFFER))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
OBJ_RELEASE(mca_pmix_native_component.cache_remote);
}
if (NULL != mca_pmix_native_component.cache_global) {
scope = PMIX_GLOBAL;
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &scope, 1, PMIX_SCOPE_T))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &mca_pmix_native_component.cache_global, 1, OPAL_BUFFER))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
OBJ_RELEASE(mca_pmix_native_component.cache_global);
}
/* create a callback object as we need to pass it to the
* recv routine so we know which callback to use when
* the return message is recvd */
cb = OBJ_NEW(pmix_cb_t);
cb->cbfunc = cbfunc;
cb->cbdata = cbdata;
/* push the message into our event base to send to the server */
PMIX_ACTIVATE_SEND_RECV(msg, fencenb_cbfunc, cb);
return OPAL_SUCCESS;
}
static int native_get(const opal_process_name_t *id,
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
const char *key,
opal_value_t **kv)
{
opal_buffer_t *msg, *bptr;
pmix_cmd_t cmd = PMIX_GET_CMD;
pmix_cb_t *cb;
int rc, ret;
int32_t cnt;
opal_list_t vals;
opal_value_t *kp;
bool found;
int handle;
char *seg_info;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native getting value for proc %s key %s",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
OPAL_NAME_PRINT(*id), key);
/* first see if we already have the info in our dstore */
OBJ_CONSTRUCT(&vals, opal_list_t);
if (1 == sm_flag) {
handle = opal_dstore_modex;
} else {
handle = opal_dstore_internal;
}
opal_proc_t *myproc = opal_proc_local_get();
if (0 == opal_compare_proc(myproc->proc_name, *id)) {
handle = opal_dstore_internal;
}
if (OPAL_SUCCESS == opal_dstore.fetch(handle, id,
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
key, &vals)) {
*kv = (opal_value_t*)opal_list_remove_first(&vals);
OPAL_LIST_DESTRUCT(&vals);
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native value retrieved from dstore",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
return OPAL_SUCCESS;
}
if (NULL == mca_pmix_native_component.uri) {
/* no server available, so just return */
return OPAL_ERR_NOT_FOUND;
}
/* nope - see if we can get it */
msg = OBJ_NEW(opal_buffer_t);
/* pack the get cmd */
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &cmd, 1, PMIX_CMD_T))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
/* pack the request information - we'll get the entire blob
* for this proc, so we don't need to pass the key */
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, id, 1, OPAL_NAME))) {
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return rc;
}
/* create a callback object as we need to pass it to the
* recv routine so we know which callback to use when
* the return message is recvd */
cb = OBJ_NEW(pmix_cb_t);
cb->active = true;
/* push the message into our event base to send to the server */
PMIX_ACTIVATE_SEND_RECV(msg, wait_cbfunc, cb);
/* wait for the data to return */
PMIX_WAIT_FOR_COMPLETION(cb->active);
/* we have received the entire data blob for this process - unpack
* and cache all values, keeping the one we requested to return
* to the caller */
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(&cb->data, &ret, &cnt, OPAL_INT))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(cb);
return rc;
}
found = false;
if (1 == sm_flag) {
opal_process_name_t uid;
unpack_segment_info(&cb->data, &uid, &seg_info);
if (NULL != seg_info) {
pmix_sm_attach(uid.jobid, seg_info);
}
OBJ_CONSTRUCT(&vals, opal_list_t);
if (OPAL_SUCCESS == opal_dstore.fetch(opal_dstore_modex, id,
key, &vals)) {
*kv = (opal_value_t*)opal_list_remove_first(&vals);
OPAL_LIST_DESTRUCT(&vals);
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native value retrieved from dstore",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
found = true;
rc = OPAL_SUCCESS;
} else {
rc = OPAL_ERROR;
}
} else {
cnt = 1;
while (OPAL_SUCCESS == (rc = opal_dss.unpack(&cb->data, &bptr, &cnt, OPAL_BUFFER))) {
while (OPAL_SUCCESS == (rc = opal_dss.unpack(bptr, &kp, &cnt, OPAL_VALUE))) {
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native retrieved %s (%s) from server for proc %s",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME), kp->key,
(OPAL_STRING == kp->type) ? kp->data.string : "NS",
OPAL_NAME_PRINT(*id));
if (OPAL_SUCCESS != (ret = opal_dstore.store(opal_dstore_internal, id, kp))) {
OPAL_ERROR_LOG(ret);
}
if (0 == strcmp(key, kp->key)) {
*kv = kp;
found = true;
} else {
OBJ_RELEASE(kp);
}
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
}
if (OPAL_ERR_UNPACK_READ_PAST_END_OF_BUFFER != rc) {
OPAL_ERROR_LOG(rc);
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
}
OBJ_RELEASE(bptr);
cnt = 1;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
}
if (OPAL_ERR_UNPACK_READ_PAST_END_OF_BUFFER != rc) {
OPAL_ERROR_LOG(rc);
} else {
rc = OPAL_SUCCESS;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
}
}
OBJ_RELEASE(cb);
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native get completed",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
if (found) {
return OPAL_SUCCESS;
}
/* we didn't find the requested data - pass back a
* status that indicates the source of the problem,
* either during the data fetch, message unpacking,
* or not found */
*kv = NULL;
if (OPAL_SUCCESS == rc) {
if (OPAL_SUCCESS == ret) {
rc = OPAL_ERR_NOT_FOUND;
} else {
rc = ret;
}
}
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
return rc;
}
static void native_get_nb(const opal_process_name_t *id,
const char *key,
opal_pmix_cbfunc_t cbfunc,
void *cbdata)
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
{
return;
}
static int native_publish(const char service_name[],
opal_list_t *info,
const char port[])
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
{
return OPAL_SUCCESS;
}
static int native_lookup(const char service_name[],
opal_list_t *info,
char port[], int portLen)
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
{
return OPAL_ERR_NOT_IMPLEMENTED;
}
static int native_unpublish(const char service_name[],
opal_list_t *info)
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
{
return OPAL_SUCCESS;;
}
static bool native_get_attr(const char *attr, opal_value_t **kv)
{
opal_buffer_t *msg, *bptr;
opal_list_t vals;
opal_value_t *kp, *lclpeers=NULL, kvn;
pmix_cmd_t cmd = PMIX_GETATTR_CMD;
char **ranks;
int rc, ret;
int32_t cnt;
bool found=false;
opal_hwloc_locality_t locality;
pmix_cb_t *cb;
uint32_t i, myrank;
opal_process_name_t id;
char *cpuset;
opal_buffer_t buf;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s pmix:native get_attr called",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
/* try to retrieve the requested value from the dstore */
OBJ_CONSTRUCT(&vals, opal_list_t);
if (OPAL_SUCCESS == opal_dstore.fetch(opal_dstore_internal, &OPAL_PROC_MY_NAME, attr, &vals)) {
*kv = (opal_value_t*)opal_list_remove_first(&vals);
OPAL_LIST_DESTRUCT(&vals);
return true;
}
if (NULL == mca_pmix_native_component.uri) {
/* no server available, so just return */
return false;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
}
/* if the value isn't yet available, then we should try to retrieve
* all the available attributes and store them for future use */
msg = OBJ_NEW(opal_buffer_t);
/* pack the cmd */
if (OPAL_SUCCESS != (rc = opal_dss.pack(msg, &cmd, 1, PMIX_CMD_T))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(msg);
return false;
}
/* create a callback object as we need to pass it to the
* recv routine so we know which callback to use when
* the return message is recvd */
cb = OBJ_NEW(pmix_cb_t);
cb->active = true;
/* push the message into our event base to send to the server */
PMIX_ACTIVATE_SEND_RECV(msg, wait_cbfunc, cb);
/* wait for the data to return */
PMIX_WAIT_FOR_COMPLETION(cb->active);
/* we have received the entire data blob for this process - unpack
* and cache all values, keeping the one we requested to return
* to the caller */
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(&cb->data, &ret, &cnt, OPAL_INT))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(cb);
return false;
}
if (OPAL_SUCCESS == ret) {
/* unpack the buffer containing the values */
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(&cb->data, &bptr, &cnt, OPAL_BUFFER))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(cb);
return false;
}
cnt = 1;
while (OPAL_SUCCESS == (rc = opal_dss.unpack(bptr, &kp, &cnt, OPAL_VALUE))) {
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s unpacked attr %s",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME), kp->key);
/* if this is the local topology, we need to save it in a special way */
#if OPAL_HAVE_HWLOC
{
hwloc_topology_t topo;
if (0 == strcmp(PMIX_LOCAL_TOPO, kp->key)) {
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s saving topology",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
/* transfer the byte object for unpacking */
OBJ_CONSTRUCT(&buf, opal_buffer_t);
opal_dss.load(&buf, kp->data.bo.bytes, kp->data.bo.size);
kp->data.bo.bytes = NULL; // protect the data region
kp->data.bo.size = 0;
OBJ_RELEASE(kp);
/* extract the topology */
cnt=1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(&buf, &topo, &cnt, OPAL_HWLOC_TOPO))) {
OPAL_ERROR_LOG(rc);
OBJ_DESTRUCT(&buf);
continue;
}
OBJ_DESTRUCT(&buf);
if (NULL == opal_hwloc_topology) {
opal_hwloc_topology = topo;
} else {
hwloc_topology_destroy(topo);
}
cnt = 1;
continue;
}
}
#endif
/* if this is the local cpuset blob, then unpack and store its contents */
if (0 == strcmp(PMIX_LOCAL_CPUSETS, kp->key)) {
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s received local cpusets",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
/* transfer the byte object for unpacking */
OBJ_CONSTRUCT(&buf, opal_buffer_t);
opal_dss.load(&buf, kp->data.bo.bytes, kp->data.bo.size);
kp->data.bo.bytes = NULL; // protect the data region
kp->data.bo.size = 0;
OBJ_RELEASE(kp);
cnt=1;
while (OPAL_SUCCESS == (rc = opal_dss.unpack(&buf, &id, &cnt, OPAL_NAME))) {
cnt=1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(&buf, &cpuset, &cnt, OPAL_STRING))) {
OPAL_ERROR_LOG(rc);
OBJ_DESTRUCT(&buf);
cnt = 1;
continue;
}
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s saving cpuset %s for local peer %s",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
(NULL == cpuset) ? "NULL" : cpuset,
OPAL_NAME_PRINT(id));
OBJ_CONSTRUCT(&kvn, opal_value_t);
kvn.key = strdup(OPAL_DSTORE_CPUSET);
kvn.type = OPAL_STRING;
kvn.data.string = cpuset;
if (OPAL_SUCCESS != (rc = opal_dstore.store(opal_dstore_internal, &id, &kvn))) {
OPAL_ERROR_LOG(rc);
OBJ_DESTRUCT(&kvn);
cnt = 1;
continue;
}
OBJ_DESTRUCT(&kvn);
}
OBJ_DESTRUCT(&buf);
if (OPAL_ERR_UNPACK_READ_PAST_END_OF_BUFFER != rc) {
OPAL_ERROR_LOG(rc);
return false;
}
cnt=1;
continue;
}
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
if (OPAL_SUCCESS != (rc = opal_dstore.store(opal_dstore_internal, &OPAL_PROC_MY_NAME, kp))) {
OPAL_ERROR_LOG(rc);
OBJ_RELEASE(kp);
cnt = 1;
continue;
}
/* save the list of local peers */
if (0 == strcmp(PMIX_LOCAL_PEERS, kp->key)) {
OBJ_RETAIN(kp);
lclpeers = kp;
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s saving local peers %s",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME), lclpeers->data.string);
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
} else if (0 == strcmp(PMIX_JOBID, kp->key)) {
native_pname.jobid = kp->data.uint32;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
} else if (0 == strcmp(PMIX_RANK, kp->key)) {
native_pname.vpid = kp->data.uint32;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
}
if (0 == strcmp(attr, kp->key)) {
OBJ_RETAIN(kp);
*kv = kp;
found = true;
}
OBJ_RELEASE(kp);
cnt = 1;
}
OBJ_RELEASE(bptr);
if (OPAL_ERR_UNPACK_READ_PAST_END_OF_BUFFER != rc) {
OPAL_ERROR_LOG(rc);
return false;
}
} else {
OPAL_ERROR_LOG(ret);
OBJ_RELEASE(cb);
return false;
}
OBJ_RELEASE(cb);
opal_proc_set_name(&native_pname);
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
/* if the list of local peers wasn't included, then we are done */
if (NULL == lclpeers) {
opal_output_verbose(0, opal_pmix_base_framework.framework_output,
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
"%s no local peers reported",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME));
return found;
}
/* baseline all the procs as nonlocal */
myrank = native_pname.vpid;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
/* we only need to set locality for each local rank as "not found"
* equates to "non local" */
ranks = opal_argv_split(lclpeers->data.string, ',');
for (i=0; NULL != ranks[i]; i++) {
2014-10-03 06:54:08 +04:00
uint32_t vid = strtoul(ranks[i], NULL, 10);
if (myrank == vid) {
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
continue;
}
native_pname.vpid = vid;
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
#if OPAL_HAVE_HWLOC
OBJ_CONSTRUCT(&vals, opal_list_t);
if (OPAL_SUCCESS != (rc = opal_dstore.fetch(opal_dstore_internal, &native_pname,
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
OPAL_DSTORE_CPUSET, &vals))) {
opal_output_verbose(2, opal_pmix_base_framework.framework_output,
"%s cpuset for local proc %s not found",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
OPAL_NAME_PRINT(native_pname));
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
OPAL_LIST_DESTRUCT(&vals);
/* even though the cpuset wasn't found, we at least know it is
* on the same node with us */
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
locality = OPAL_PROC_ON_CLUSTER | OPAL_PROC_ON_CU | OPAL_PROC_ON_NODE;
} else {
kp = (opal_value_t*)opal_list_get_first(&vals);
if (NULL == kp->data.string) {
/* if we share a node, but we don't know anything more, then
* mark us as on the node as this is all we know
*/
locality = OPAL_PROC_ON_CLUSTER | OPAL_PROC_ON_CU | OPAL_PROC_ON_NODE;
} else {
/* determine relative location on our node */
locality = opal_hwloc_base_get_relative_locality(opal_hwloc_topology,
opal_process_info.cpuset,
kp->data.string);
}
OPAL_LIST_DESTRUCT(&vals);
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
}
#else
/* all we know is we share a node */
locality = OPAL_PROC_ON_CLUSTER | OPAL_PROC_ON_CU | OPAL_PROC_ON_NODE;
#endif
OPAL_OUTPUT_VERBOSE((1, opal_pmix_base_framework.framework_output,
"%s pmix:native proc %s locality %s",
OPAL_NAME_PRINT(OPAL_PROC_MY_NAME),
OPAL_NAME_PRINT(native_pname),
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
opal_hwloc_base_print_locality(locality)));
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
OBJ_CONSTRUCT(&kvn, opal_value_t);
kvn.key = strdup(OPAL_DSTORE_LOCALITY);
kvn.type = OPAL_UINT16;
kvn.data.uint16 = locality;
(void)opal_dstore.store(opal_dstore_internal, &native_pname, &kvn);
Per the PMIx RFC: WHAT: Merge the PMIx branch into the devel repo, creating a new OPAL “lmix” framework to abstract PMI support for all RTEs. Replace the ORTE daemon-level collectives with a new PMIx server and update the ORTE grpcomm framework to support server-to-server collectives WHY: We’ve had problems dealing with variations in PMI implementations, and need to extend the existing PMI definitions to meet exascale requirements. WHEN: Mon, Aug 25 WHERE: https://github.com/rhc54/ompi-svn-mirror.git Several community members have been working on a refactoring of the current PMI support within OMPI. Although the APIs are common, Slurm and Cray implement a different range of capabilities, and package them differently. For example, Cray provides an integrated PMI-1/2 library, while Slurm separates the two and requires the user to specify the one to be used at runtime. In addition, several bugs in the Slurm implementations have caused problems requiring extra coding. All this has led to a slew of #if’s in the PMI code and bugs when the corner-case logic for one implementation accidentally traps the other. Extending this support to other implementations would have increased this complexity to an unacceptable level. Accordingly, we have: * created a new OPAL “pmix” framework to abstract the PMI support, with separate components for Cray, Slurm PMI-1, and Slurm PMI-2 implementations. * Replaced the current ORTE grpcomm daemon-based collective operation with an integrated PMIx server, and updated the grpcomm APIs to provide more flexible, multi-algorithm support for collective operations. At this time, only the xcast and allgather operations are supported. * Replaced the current global collective id with a signature based on the names of the participating procs. The allows an unlimited number of collectives to be executed by any group of processes, subject to the requirement that only one collective can be active at a time for a unique combination of procs. Note that a proc can be involved in any number of simultaneous collectives - it is the specific combination of procs that is subject to the constraint * removed the prior OMPI/OPAL modex code * added new macros for executing modex send/recv to simplify use of the new APIs. The send macros allow the caller to specify whether or not the BTL supports async modex operations - if so, then the non-blocking “fence” operation is used, if the active PMIx component supports it. Otherwise, the default is a full blocking modex exchange as we currently perform. * retained the current flag that directs us to use a blocking fence operation, but only to retrieve data upon demand This commit was SVN r32570.
2014-08-21 22:56:47 +04:00
OBJ_DESTRUCT(&kvn);
}
opal_argv_free(ranks);
return found;
}
static int native_get_attr_nb(const char *attr,
opal_pmix_cbfunc_t cbfunc,
void *cbdata)
{
return OPAL_ERR_NOT_IMPLEMENTED;
}
static int native_job_connect(const char jobId[])
{
return OPAL_ERR_NOT_IMPLEMENTED;
}
static int native_job_disconnect(const char jobId[])
{
return OPAL_ERR_NOT_IMPLEMENTED;
}
/*** INSTANTIATE INTERNAL CLASSES ***/
static void scon(pmix_usock_send_t *p)
{
p->hdr.type = 0;
p->hdr.tag = UINT32_MAX;
p->hdr.nbytes = 0;
p->data = NULL;
p->hdr_sent = false;
p->sdptr = NULL;
p->sdbytes = 0;
}
OBJ_CLASS_INSTANCE(pmix_usock_send_t,
opal_list_item_t,
scon, NULL);
static void rcon(pmix_usock_recv_t *p)
{
p->hdr.type = 0;
p->hdr.tag = UINT32_MAX;
p->hdr.nbytes = 0;
p->data = NULL;
p->hdr_recvd = false;
p->rdptr = NULL;
p->rdbytes = 0;
}
OBJ_CLASS_INSTANCE(pmix_usock_recv_t,
opal_list_item_t,
rcon, NULL);
static void prcon(pmix_usock_posted_recv_t *p)
{
p->tag = UINT32_MAX;
p->cbfunc = NULL;
p->cbdata = NULL;
}
OBJ_CLASS_INSTANCE(pmix_usock_posted_recv_t,
opal_list_item_t,
prcon, NULL);
static void cbcon(pmix_cb_t *p)
{
p->active = false;
OBJ_CONSTRUCT(&p->data, opal_buffer_t);
p->cbfunc = NULL;
p->cbdata = NULL;
}
static void cbdes(pmix_cb_t *p)
{
OBJ_DESTRUCT(&p->data);
}
OBJ_CLASS_INSTANCE(pmix_cb_t,
opal_object_t,
cbcon, cbdes);
static void srcon(pmix_usock_sr_t *p)
{
p->bfr = NULL;
p->cbfunc = NULL;
p->cbdata = NULL;
}
OBJ_CLASS_INSTANCE(pmix_usock_sr_t,
opal_object_t,
srcon, NULL);