ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
bb65835816
openmpi/ompi/proc/proc.c
Ralph Castain aec5cd08bd 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 18:56:47 +00:00

806 строки
30 KiB
C
Исходник Ответственный История

/*
* Copyright (c) 2004-2006 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2011 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2006 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2006 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2006-2014 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2012 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2013-2014 Intel, Inc. All rights reserved
* Copyright (c) 2014 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include <string.h>
#include <strings.h>
#include "ompi/constants.h"
#include "opal/datatype/opal_convertor.h"
#include "opal/threads/mutex.h"
#include "opal/dss/dss.h"
#include "opal/util/arch.h"
#include "opal/util/show_help.h"
#include "opal/mca/dstore/dstore.h"
#include "opal/mca/hwloc/base/base.h"
#include "opal/mca/pmix/pmix.h"
#include "ompi/proc/proc.h"
#include "ompi/datatype/ompi_datatype.h"
#include "ompi/runtime/mpiruntime.h"
#include "ompi/runtime/params.h"
static opal_list_t ompi_proc_list;
static opal_mutex_t ompi_proc_lock;
ompi_proc_t* ompi_proc_local_proc = NULL;
static void ompi_proc_construct(ompi_proc_t* proc);
static void ompi_proc_destruct(ompi_proc_t* proc);
OBJ_CLASS_INSTANCE(
ompi_proc_t,
opal_proc_t,
ompi_proc_construct,
ompi_proc_destruct
);
void ompi_proc_construct(ompi_proc_t* proc)
{
bzero(proc->proc_endpoints, sizeof(proc->proc_endpoints));
/* By default all processors are supposedly having the same architecture as me. Thus,
* by default we run in a homogeneous environment. Later, when the RTE can tell us
* the arch of the remote nodes, we will have to set the convertors to the correct
* architecture.
*/
OBJ_RETAIN( ompi_mpi_local_convertor );
proc->super.proc_convertor = ompi_mpi_local_convertor;
}
void ompi_proc_destruct(ompi_proc_t* proc)
{
/* As all the convertors are created with OBJ_NEW we can just call OBJ_RELEASE. All, except
* the local convertor, will get destroyed at some point here. If the reference count is correct
* the local convertor (who has the reference count increased in the datatype) will not get
* destroyed here. It will be destroyed later when the ompi_datatype_finalize is called.
*/
OBJ_RELEASE( proc->super.proc_convertor );
if (NULL != proc->super.proc_hostname) {
free(proc->super.proc_hostname);
}
OPAL_THREAD_LOCK(&ompi_proc_lock);
opal_list_remove_item(&ompi_proc_list, (opal_list_item_t*)proc);
OPAL_THREAD_UNLOCK(&ompi_proc_lock);
}
int ompi_proc_init(void)
{
ompi_vpid_t i;
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
int ret;
#endif
OBJ_CONSTRUCT(&ompi_proc_list, opal_list_t);
OBJ_CONSTRUCT(&ompi_proc_lock, opal_mutex_t);
/* create proc structures and find self */
for( i = 0; i < ompi_process_info.num_procs; i++ ) {
ompi_proc_t *proc = OBJ_NEW(ompi_proc_t);
opal_list_append(&ompi_proc_list, (opal_list_item_t*)proc);
OMPI_CAST_RTE_NAME(&proc->super.proc_name)->jobid = OMPI_PROC_MY_NAME->jobid;
OMPI_CAST_RTE_NAME(&proc->super.proc_name)->vpid = i;
if (i == OMPI_PROC_MY_NAME->vpid) {
ompi_proc_local_proc = proc;
proc->super.proc_flags = OPAL_PROC_ALL_LOCAL;
proc->super.proc_hostname = strdup(ompi_process_info.nodename);
proc->super.proc_arch = opal_local_arch;
/* Register the local proc with OPAL */
opal_proc_local_set(&proc->super);
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
/* add our arch to the modex */
OPAL_MODEX_SEND_STRING(ret, PMIX_SYNC_REQD, PMIX_REMOTE, OPAL_DSTORE_ARCH,
&proc->super.proc_arch, OPAL_UINT32);
if (OPAL_SUCCESS != ret) {
return ret;
}
#endif
}
}
return OMPI_SUCCESS;
}
/**
* The process creation is split into two steps. The second step
* is the important one, it sets the properties of the remote
* process, such as architecture, node name and locality flags.
*
* This function is to be called __only__ after the modex exchange
* has been performed, in order to allow the modex to carry the data
* instead of requiring the runtime to provide it.
*/
int ompi_proc_complete_init(void)
{
ompi_proc_t *proc;
int ret, errcode = OMPI_SUCCESS;
opal_list_t myvals;
opal_value_t *kv;
OPAL_THREAD_LOCK(&ompi_proc_lock);
OPAL_LIST_FOREACH(proc, &ompi_proc_list, ompi_proc_t) {
if (OMPI_CAST_RTE_NAME(&proc->super.proc_name)->vpid != OMPI_PROC_MY_NAME->vpid) {
/* get the locality information - do not use modex recv for
* this request as that will automatically cause the hostname
* to be loaded as well. All RTEs are required to provide this
* information at startup for procs on our node. Thus, not
* finding the info indicates that the proc is non-local.
*/
OBJ_CONSTRUCT(&myvals, opal_list_t);
if (OMPI_SUCCESS != (ret = opal_dstore.fetch(opal_dstore_internal,
(opal_identifier_t*)&proc->super.proc_name,
OPAL_DSTORE_LOCALITY, &myvals))) {
proc->super.proc_flags = OPAL_PROC_NON_LOCAL;
} else {
kv = (opal_value_t*)opal_list_get_first(&myvals);
proc->super.proc_flags = kv->data.uint16;
}
OPAL_LIST_DESTRUCT(&myvals);
if (ompi_process_info.num_procs < ompi_hostname_cutoff) {
/* IF the number of procs falls below the specified cutoff,
* then we assume the job is small enough that retrieving
* the hostname (which will typically cause retrieval of
* ALL modex info for this proc) will have no appreciable
* impact on launch scaling
*/
OPAL_MODEX_RECV_VALUE(ret, OPAL_DSTORE_HOSTNAME, (opal_proc_t*)&proc->super,
(char**)&(proc->super.proc_hostname), OPAL_STRING);
if (OPAL_SUCCESS != ret) {
errcode = ret;
break;
}
} else {
/* just set the hostname to NULL for now - we'll fill it in
* as modex_recv's are called for procs we will talk to, thus
* avoiding retrieval of ALL modex info for this proc until
* required. Transports that delay calling modex_recv until
* first message will therefore scale better than those that
* call modex_recv on all procs during init.
*/
proc->super.proc_hostname = NULL;
}
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
/* get the remote architecture - this might force a modex except
* for those environments where the RM provides it */
{
uint32_t *ui32ptr;
ui32ptr = &(proc->super.proc_arch);
OPAL_MODEX_RECV_VALUE(ret, OPAL_DSTORE_ARCH, (opal_proc_t*)&proc->super,
(void**)&ui32ptr, OPAL_UINT32);
if (OPAL_SUCCESS == ret) {
/* if arch is different than mine, create a new convertor for this proc */
if (proc->super.proc_arch != opal_local_arch) {
OBJ_RELEASE(proc->super.proc_convertor);
proc->super.proc_convertor = opal_convertor_create(proc->super.proc_arch, 0);
}
} else if (OMPI_ERR_NOT_IMPLEMENTED == ret) {
proc->super.proc_arch = opal_local_arch;
} else {
errcode = ret;
break;
}
}
#else
/* must be same arch as my own */
proc->super.proc_arch = opal_local_arch;
#endif
}
}
OPAL_THREAD_UNLOCK(&ompi_proc_lock);
return errcode;
}
int ompi_proc_finalize (void)
{
opal_list_item_t *item;
/* Unregister the local proc from OPAL */
opal_proc_local_set(NULL);
/* remove all items from list and destroy them. Since we cannot know
* the reference count of the procs for certain, it is possible that
* a single OBJ_RELEASE won't drive the count to zero, and hence will
* not release the memory. Accordingly, we cycle through the list here,
* calling release on each item.
*
* This will cycle until it forces the reference count of each item
* to zero, thus causing the destructor to run - which will remove
* the item from the list!
*
* We cannot do this under the thread lock as the destructor will
* call it when removing the item from the list. However, this function
* is ONLY called from MPI_Finalize, and all threads are prohibited from
* calling an MPI function once ANY thread has called MPI_Finalize. Of
* course, multiple threads are allowed to call MPI_Finalize, so this
* function may get called multiple times by various threads. We believe
* it is thread safe to do so...though it may not -appear- to be so
* without walking through the entire list/destructor sequence.
*/
while (opal_list_get_end(&ompi_proc_list) != (item = opal_list_get_first(&ompi_proc_list))) {
OBJ_RELEASE(item);
}
/* now destruct the list and thread lock */
OBJ_DESTRUCT(&ompi_proc_list);
OBJ_DESTRUCT(&ompi_proc_lock);
return OMPI_SUCCESS;
}
ompi_proc_t** ompi_proc_world(size_t *size)
{
ompi_proc_t **procs;
ompi_proc_t *proc;
size_t count = 0;
ompi_rte_cmp_bitmask_t mask;
ompi_process_name_t my_name;
/* check bozo case */
if (NULL == ompi_proc_local_proc) {
return NULL;
}
mask = OMPI_RTE_CMP_JOBID;
my_name = *OMPI_CAST_RTE_NAME(&ompi_proc_local_proc->super.proc_name);
/* First count how many match this jobid */
OPAL_THREAD_LOCK(&ompi_proc_lock);
for (proc = (ompi_proc_t*)opal_list_get_first(&ompi_proc_list);
proc != (ompi_proc_t*)opal_list_get_end(&ompi_proc_list);
proc = (ompi_proc_t*)opal_list_get_next(proc)) {
if (OPAL_EQUAL == ompi_rte_compare_name_fields(mask, OMPI_CAST_RTE_NAME(&proc->super.proc_name), &my_name)) {
++count;
}
}
/* allocate an array */
procs = (ompi_proc_t**) malloc(count * sizeof(ompi_proc_t*));
if (NULL == procs) {
OPAL_THREAD_UNLOCK(&ompi_proc_lock);
return NULL;
}
/* now save only the procs that match this jobid */
count = 0;
for (proc = (ompi_proc_t*)opal_list_get_first(&ompi_proc_list);
proc != (ompi_proc_t*)opal_list_get_end(&ompi_proc_list);
proc = (ompi_proc_t*)opal_list_get_next(proc)) {
if (OPAL_EQUAL == ompi_rte_compare_name_fields(mask, &proc->super.proc_name, &my_name)) {
/* DO NOT RETAIN THIS OBJECT - the reference count on this
* object will be adjusted by external callers. The intent
* here is to allow the reference count to drop to zero if
* the app no longer desires to communicate with this proc.
* For example, the proc may call comm_disconnect on all
* communicators involving this proc. In such cases, we want
* the proc object to be removed from the list. By not incrementing
* the reference count here, we allow this to occur.
*
* We don't implement that yet, but we are still safe for now as
* the OBJ_NEW in ompi_proc_init owns the initial reference
* count which cannot be released until ompi_proc_finalize is
* called.
*/
procs[count++] = proc;
}
}
OPAL_THREAD_UNLOCK(&ompi_proc_lock);
*size = count;
return procs;
}
ompi_proc_t** ompi_proc_all(size_t* size)
{
ompi_proc_t **procs =
(ompi_proc_t**) malloc(opal_list_get_size(&ompi_proc_list) * sizeof(ompi_proc_t*));
ompi_proc_t *proc;
size_t count = 0;
if (NULL == procs) {
return NULL;
}
OPAL_THREAD_LOCK(&ompi_proc_lock);
for(proc = (ompi_proc_t*)opal_list_get_first(&ompi_proc_list);
proc != (ompi_proc_t*)opal_list_get_end(&ompi_proc_list);
proc = (ompi_proc_t*)opal_list_get_next(proc)) {
/* We know this isn't consistent with the behavior in ompi_proc_world,
* but we are leaving the RETAIN for now because the code using this function
* assumes that the results need to be released when done. It will
* be cleaned up later as the "fix" will impact other places in
* the code
*/
OBJ_RETAIN(proc);
procs[count++] = proc;
}
OPAL_THREAD_UNLOCK(&ompi_proc_lock);
*size = count;
return procs;
}
ompi_proc_t** ompi_proc_self(size_t* size)
{
ompi_proc_t **procs = (ompi_proc_t**) malloc(sizeof(ompi_proc_t*));
if (NULL == procs) {
return NULL;
}
/* We know this isn't consistent with the behavior in ompi_proc_world,
* but we are leaving the RETAIN for now because the code using this function
* assumes that the results need to be released when done. It will
* be cleaned up later as the "fix" will impact other places in
* the code
*/
OBJ_RETAIN(ompi_proc_local_proc);
*procs = ompi_proc_local_proc;
*size = 1;
return procs;
}
ompi_proc_t * ompi_proc_find ( const ompi_process_name_t * name )
{
ompi_proc_t *proc, *rproc=NULL;
ompi_rte_cmp_bitmask_t mask;
/* return the proc-struct which matches this jobid+process id */
mask = OMPI_RTE_CMP_JOBID | OMPI_RTE_CMP_VPID;
OPAL_THREAD_LOCK(&ompi_proc_lock);
for(proc = (ompi_proc_t*)opal_list_get_first(&ompi_proc_list);
proc != (ompi_proc_t*)opal_list_get_end(&ompi_proc_list);
proc = (ompi_proc_t*)opal_list_get_next(proc)) {
if (OPAL_EQUAL == ompi_rte_compare_name_fields(mask, &proc->super.proc_name, name)) {
rproc = proc;
break;
}
}
OPAL_THREAD_UNLOCK(&ompi_proc_lock);
return rproc;
}
int ompi_proc_refresh(void)
{
ompi_proc_t *proc = NULL;
opal_list_item_t *item = NULL;
ompi_vpid_t i = 0;
int ret=OMPI_SUCCESS;
opal_list_t myvals;
opal_value_t *kv;
OPAL_THREAD_LOCK(&ompi_proc_lock);
for( item = opal_list_get_first(&ompi_proc_list), i = 0;
item != opal_list_get_end(&ompi_proc_list);
item = opal_list_get_next(item), ++i ) {
proc = (ompi_proc_t*)item;
/* Does not change: proc->super.proc_name.vpid */
OMPI_CAST_RTE_NAME(&proc->super.proc_name)->jobid = OMPI_PROC_MY_NAME->jobid;
/* Make sure to clear the local flag before we set it below */
proc->super.proc_flags = 0;
if (i == OMPI_PROC_MY_NAME->vpid) {
ompi_proc_local_proc = proc;
proc->super.proc_flags = OPAL_PROC_ALL_LOCAL;
proc->super.proc_hostname = ompi_process_info.nodename;
proc->super.proc_arch = opal_local_arch;
opal_proc_local_set(&proc->super);
} else {
/* get the locality information - do not use modex recv for
* this request as that will automatically cause the hostname
* to be loaded as well. All RTEs are required to provide this
* information at startup for procs on our node. Thus, not
* finding the info indicates that the proc is non-local.
*/
OBJ_CONSTRUCT(&myvals, opal_list_t);
if (OMPI_SUCCESS != (ret = opal_dstore.fetch(opal_dstore_internal,
(opal_identifier_t*)&proc->super.proc_name,
OPAL_DSTORE_LOCALITY, &myvals))) {
proc->super.proc_flags = OPAL_PROC_NON_LOCAL;
} else {
kv = (opal_value_t*)opal_list_get_first(&myvals);
proc->super.proc_flags = kv->data.uint16;
}
OPAL_LIST_DESTRUCT(&myvals);
if (ompi_process_info.num_procs < ompi_hostname_cutoff) {
/* IF the number of procs falls below the specified cutoff,
* then we assume the job is small enough that retrieving
* the hostname (which will typically cause retrieval of
* ALL modex info for this proc) will have no appreciable
* impact on launch scaling
*/
OPAL_MODEX_RECV_VALUE(ret, OPAL_DSTORE_HOSTNAME, (opal_proc_t*)&proc->super,
(char**)&(proc->super.proc_hostname), OPAL_STRING);
if (OMPI_SUCCESS != ret) {
break;
}
} else {
/* just set the hostname to NULL for now - we'll fill it in
* as modex_recv's are called for procs we will talk to, thus
* avoiding retrieval of ALL modex info for this proc until
* required. Transports that delay calling modex_recv until
* first message will therefore scale better than those that
* call modex_recv on all procs during init.
*/
proc->super.proc_hostname = NULL;
}
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
{
/* get the remote architecture */
uint32_t* uiptr = &(proc->super.proc_arch);
OPAL_MODEX_RECV_VALUE(ret, OPAL_DSTORE_ARCH, (opal_proc_t*)&proc->super,
(void**)&uiptr, OPAL_UINT32);
if (OMPI_SUCCESS != ret) {
break;
}
/* if arch is different than mine, create a new convertor for this proc */
if (proc->super.proc_arch != opal_local_arch) {
OBJ_RELEASE(proc->super.proc_convertor);
proc->super.proc_convertor = opal_convertor_create(proc->super.proc_arch, 0);
}
}
#else
/* must be same arch as my own */
proc->super.proc_arch = opal_local_arch;
#endif
}
}
OPAL_THREAD_UNLOCK(&ompi_proc_lock);
return ret;
}
int
ompi_proc_pack(ompi_proc_t **proclist, int proclistsize,
bool full_info,
opal_buffer_t* buf)
{
int i, rc;
OPAL_THREAD_LOCK(&ompi_proc_lock);
/* cycle through the provided array, packing the OMPI level
* data for each proc. This data may or may not be included
* in any subsequent modex operation, so we include it here
* to ensure completion of a connect/accept handshake. See
* the ompi/mca/dpm framework for an example of where and how
* this info is used.
*
* Eventually, we will review the procedures that call this
* function to see if duplication of communication can be
* reduced. For now, just go ahead and pack the info so it
* can be sent.
*/
for (i=0; i<proclistsize; i++) {
rc = opal_dss.pack(buf, &(proclist[i]->super.proc_name), 1, OMPI_NAME);
if(rc != OPAL_SUCCESS) {
OMPI_ERROR_LOG(rc);
OPAL_THREAD_UNLOCK(&ompi_proc_lock);
return rc;
}
if (full_info) {
int32_t num_entries;
opal_value_t *kv;
opal_list_t data;
/* fetch all info we know about the peer - while
* the remote procs may already know some of it, we cannot
* be certain they do. So we must include a full dump of
* everything we know about this proc
*/
OBJ_CONSTRUCT(&data, opal_list_t);
rc = opal_dstore.fetch(opal_dstore_internal,
(opal_identifier_t*)&proclist[i]->super.proc_name,
NULL, &data);
if (OPAL_SUCCESS != rc) {
OMPI_ERROR_LOG(rc);
num_entries = 0;
} else {
/* count the number of entries we will send */
num_entries = opal_list_get_size(&data);
}
/* put the number of entries into the buffer */
rc = opal_dss.pack(buf, &num_entries, 1, OPAL_INT32);
if (OPAL_SUCCESS != rc) {
OMPI_ERROR_LOG(rc);
break;
}
/* if there are entries, store them */
while (NULL != (kv = (opal_value_t*)opal_list_remove_first(&data))) {
if (OPAL_SUCCESS != (rc = opal_dss.pack(buf, &kv, 1, OPAL_VALUE))) {
OMPI_ERROR_LOG(rc);
break;
}
OBJ_RELEASE(kv);
}
OBJ_DESTRUCT(&data);
} else {
rc = opal_dss.pack(buf, &(proclist[i]->super.proc_arch), 1, OPAL_UINT32);
if(rc != OPAL_SUCCESS) {
OMPI_ERROR_LOG(rc);
OPAL_THREAD_UNLOCK(&ompi_proc_lock);
return rc;
}
rc = opal_dss.pack(buf, &(proclist[i]->super.proc_hostname), 1, OPAL_STRING);
if(rc != OPAL_SUCCESS) {
OMPI_ERROR_LOG(rc);
OPAL_THREAD_UNLOCK(&ompi_proc_lock);
return rc;
}
}
}
OPAL_THREAD_UNLOCK(&ompi_proc_lock);
return OMPI_SUCCESS;
}
static ompi_proc_t *
ompi_proc_find_and_add(const ompi_process_name_t * name, bool* isnew)
{
ompi_proc_t *proc, *rproc = NULL;
ompi_rte_cmp_bitmask_t mask;
/* return the proc-struct which matches this jobid+process id */
mask = OMPI_RTE_CMP_JOBID | OMPI_RTE_CMP_VPID;
OPAL_THREAD_LOCK(&ompi_proc_lock);
for(proc = (ompi_proc_t*)opal_list_get_first(&ompi_proc_list);
proc != (ompi_proc_t*)opal_list_get_end(&ompi_proc_list);
proc = (ompi_proc_t*)opal_list_get_next(proc)) {
if (OPAL_EQUAL == ompi_rte_compare_name_fields(mask, &proc->super.proc_name, name)) {
rproc = proc;
*isnew = false;
break;
}
}
/* if we didn't find this proc in the list, create a new
* proc_t and append it to the list
*/
if (NULL == rproc) {
*isnew = true;
rproc = OBJ_NEW(ompi_proc_t);
if (NULL != rproc) {
opal_list_append(&ompi_proc_list, (opal_list_item_t*)rproc);
*OMPI_CAST_RTE_NAME(&rproc->super.proc_name) = *name;
}
/* caller had better fill in the rest of the proc, or there's
going to be pain later... */
}
OPAL_THREAD_UNLOCK(&ompi_proc_lock);
return rproc;
}
int
ompi_proc_unpack(opal_buffer_t* buf,
int proclistsize, ompi_proc_t ***proclist,
bool full_info,
int *newproclistsize, ompi_proc_t ***newproclist)
{
int i;
size_t newprocs_len = 0;
ompi_proc_t **plist=NULL, **newprocs = NULL;
opal_list_t myvals;
opal_value_t *kv;
/* do not free plist *ever*, since it is used in the remote group
structure of a communicator */
plist = (ompi_proc_t **) calloc (proclistsize, sizeof (ompi_proc_t *));
if ( NULL == plist ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* free this on the way out */
newprocs = (ompi_proc_t **) calloc (proclistsize, sizeof (ompi_proc_t *));
if (NULL == newprocs) {
free(plist);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* cycle through the array of provided procs and unpack
* their info - as packed by ompi_proc_pack
*/
for ( i=0; i<proclistsize; i++ ){
int32_t count=1;
ompi_process_name_t new_name;
uint32_t new_arch;
char *new_hostname;
bool isnew = false;
int rc;
rc = opal_dss.unpack(buf, &new_name, &count, OMPI_NAME);
if (rc != OPAL_SUCCESS) {
OMPI_ERROR_LOG(rc);
free(plist);
free(newprocs);
return rc;
}
if (!full_info) {
rc = opal_dss.unpack(buf, &new_arch, &count, OPAL_UINT32);
if (rc != OPAL_SUCCESS) {
OMPI_ERROR_LOG(rc);
free(plist);
free(newprocs);
return rc;
}
rc = opal_dss.unpack(buf, &new_hostname, &count, OPAL_STRING);
if (rc != OPAL_SUCCESS) {
OMPI_ERROR_LOG(rc);
free(plist);
free(newprocs);
return rc;
}
}
/* see if this proc is already on our ompi_proc_list */
plist[i] = ompi_proc_find_and_add(&new_name, &isnew);
if (isnew) {
/* if not, then it was added, so update the values
* in the proc_t struct with the info that was passed
* to us
*/
newprocs[newprocs_len++] = plist[i];
if (full_info) {
int32_t num_recvd_entries;
int32_t cnt;
int32_t j;
/* unpack the number of entries for this proc */
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(buf, &num_recvd_entries, &cnt, OPAL_INT32))) {
OMPI_ERROR_LOG(rc);
break;
}
/*
* Extract the attribute names and values
*/
for (j = 0; j < num_recvd_entries; j++) {
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(buf, &kv, &cnt, OPAL_VALUE))) {
OMPI_ERROR_LOG(rc);
break;
}
/* if this is me, ignore the data - we already have it in the db */
if (OPAL_EQUAL != ompi_rte_compare_name_fields(OMPI_RTE_CMP_ALL,
OMPI_PROC_MY_NAME, &new_name)) {
/* store it in the database */
if (OPAL_SUCCESS != (rc = opal_dstore.store(opal_dstore_internal,
(opal_identifier_t*)&new_name, kv))) {
OMPI_ERROR_LOG(rc);
}
}
OBJ_RELEASE(kv);
}
/* RHC: compute locality */
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
OBJ_CONSTRUCT(&myvals, opal_list_t);
rc = opal_dstore.fetch(opal_dstore_internal,
(opal_identifier_t*)&new_name,
OPAL_DSTORE_ARCH, &myvals);
if( OPAL_SUCCESS == rc ) {
kv = (opal_value_t*)opal_list_get_first(&myvals);
new_arch = kv->data.uint32;
} else {
new_arch = opal_local_arch;
}
OPAL_LIST_DESTRUCT(&myvals);
#else
new_arch = opal_local_arch;
#endif
if (ompi_process_info.num_procs < ompi_hostname_cutoff) {
/* retrieve the hostname */
OBJ_CONSTRUCT(&myvals, opal_list_t);
rc = opal_dstore.fetch(opal_dstore_internal,
(opal_identifier_t*)&new_name,
OPAL_DSTORE_HOSTNAME, &myvals);
if( OPAL_SUCCESS == rc ) {
kv = (opal_value_t*)opal_list_get_first(&myvals);
new_hostname = strdup(kv->data.string);
} else {
new_hostname = NULL;
}
OPAL_LIST_DESTRUCT(&myvals);
} else {
/* just set the hostname to NULL for now - we'll fill it in
* as modex_recv's are called for procs we will talk to
*/
new_hostname = NULL;
}
}
/* update all the values */
plist[i]->super.proc_arch = new_arch;
/* if arch is different than mine, create a new convertor for this proc */
if (plist[i]->super.proc_arch != opal_local_arch) {
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
OBJ_RELEASE(plist[i]->super.proc_convertor);
plist[i]->super.proc_convertor = opal_convertor_create(plist[i]->super.proc_arch, 0);
#else
opal_show_help("help-mpi-runtime.txt",
"heterogeneous-support-unavailable",
true, ompi_process_info.nodename,
new_hostname == NULL ? "<hostname unavailable>" :
new_hostname);
free(plist);
free(newprocs);
return OMPI_ERR_NOT_SUPPORTED;
#endif
}
if (0 == strcmp(ompi_proc_local_proc->super.proc_hostname, new_hostname)) {
plist[i]->super.proc_flags |= (OPAL_PROC_ON_NODE | OPAL_PROC_ON_CU | OPAL_PROC_ON_CLUSTER);
}
/* Save the hostname */
plist[i]->super.proc_hostname = new_hostname;
} else {
if (full_info) {
int32_t num_recvd_entries;
int32_t j, cnt;
/* discard all keys: they are already locally known */
cnt = 1;
if (OPAL_SUCCESS == (rc = opal_dss.unpack(buf, &num_recvd_entries, &cnt, OPAL_INT32))) {
for (j = 0; j < num_recvd_entries; j++) {
opal_value_t *kv;
cnt = 1;
if (OPAL_SUCCESS != (rc = opal_dss.unpack(buf, &kv, &cnt, OPAL_VALUE))) {
OMPI_ERROR_LOG(rc);
continue;
}
OBJ_RELEASE(kv);
}
} else {
OMPI_ERROR_LOG(rc);
}
}
}
}
if (NULL != newproclistsize) *newproclistsize = newprocs_len;
if (NULL != newproclist) {
*newproclist = newprocs;
} else if (newprocs != NULL) {
free(newprocs);
}
*proclist = plist;
return OMPI_SUCCESS;
}
Показать git blame Копировать постоянную ссылку