ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
cef185c8b3
openmpi/ompi/communicator/comm_init.c
Xi Luo e65fa4ff5c Bring ADAPT collective to 4.1 
This is a meta commit, that encapsulate all the ADAPT commits in the master
into a single PR for 4.1. The master commits included here are:
fe73586, a4be3bb, d712645, c2970a3, e59bde9, ee592f3 and c98e387.

Here is a detailed list of added capabilities:
* coll/adapt: Fix naming conventions and C11 atomic use
* coll/adapt: Remove unused component field in module
* Consistent handling of zero counts in the MPI API.
* Correctly handle non-blocking collectives tags
  * As it is possible to have multiple outstanding non-blocking collectives
    provided by different collective modules, we need a consistent
    mechanism to allow them to select unique tags for each instance of a
    collective.
* Add support for fallback to previous coll module on non-commutative operations (#30)
* Replace mutexes by atomic operations.
* Use the correct nbc request type (for both ibcast and ireduce)
  * coll/base: document type casts in ompi_coll_base_retain_*
* add module-wide topology cache
* use standard instead of synchronous send and add mca parameter to control mode of initial send in ireduce/ibcast
* reduce number of memory allocations
* call the default request completion.
  * Remove the requests from the Fortran lookup conversion tables before completing
    and free it.
* piggybacking Bull functionalities

Signed-off-by: Xi Luo <xluo12@vols.utk.edu>
Signed-off-by: George Bosilca <bosilca@icl.utk.edu>
Signed-off-by: Marc Sergent <marc.sergent@atos.net>
Co-authored-by: Joseph Schuchart <schuchart@hlrs.de>
Co-authored-by: Lemarinier, Pierre <pierre.lemarinier@atos.net>
Co-authored-by: pierrele <31764860+pierrele@users.noreply.github.com>
2020-09-23 11:45:45 -04:00

510 строки
21 KiB
C
Исходник Ответственный История

/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2020 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* Copyright (c) 2006-2017 University of Houston. All rights reserved.
* Copyright (c) 2007-2012 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2009 Sun Microsystems, Inc. All rights reserved.
* Copyright (c) 2012-2015 Los Alamos National Security, LLC.
* All rights reserved.
* Copyright (c) 2011-2013 Inria. All rights reserved.
* Copyright (c) 2011-2013 Universite Bordeaux 1
* All rights reserved.
* Copyright (c) 2015-2017 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* Copyright (c) 2015-2017 Intel, Inc. All rights reserved.
* Copyright (c) 2016-2017 IBM Corporation. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include <stdio.h>
#include "opal/util/bit_ops.h"
#include "opal/util/info_subscriber.h"
#include "opal/mca/pmix/pmix.h"
#include "ompi/constants.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/mca/coll/base/base.h"
#include "ompi/mca/coll/base/coll_tags.h"
#include "ompi/mca/topo/base/base.h"
#include "ompi/runtime/params.h"
#include "ompi/communicator/communicator.h"
#include "ompi/attribute/attribute.h"
#include "ompi/dpm/dpm.h"
#include "ompi/memchecker.h"
/*
** Table for Fortran <-> C communicator handle conversion
** Also used by P2P code to lookup communicator based
** on cid.
**
*/
opal_pointer_array_t ompi_mpi_communicators = {{0}};
opal_pointer_array_t ompi_comm_f_to_c_table = {{0}};
ompi_predefined_communicator_t ompi_mpi_comm_world = {{{{0}}}};
ompi_predefined_communicator_t ompi_mpi_comm_self = {{{{0}}}};
ompi_predefined_communicator_t ompi_mpi_comm_null = {{{{0}}}};
ompi_communicator_t *ompi_mpi_comm_parent = NULL;
ompi_predefined_communicator_t *ompi_mpi_comm_world_addr =
&ompi_mpi_comm_world;
ompi_predefined_communicator_t *ompi_mpi_comm_self_addr =
&ompi_mpi_comm_self;
ompi_predefined_communicator_t *ompi_mpi_comm_null_addr =
&ompi_mpi_comm_null;
static void ompi_comm_construct(ompi_communicator_t* comm);
static void ompi_comm_destruct(ompi_communicator_t* comm);
OBJ_CLASS_INSTANCE(ompi_communicator_t, opal_infosubscriber_t,
ompi_comm_construct,
ompi_comm_destruct);
/* This is the counter for the number of communicators, which contain
process with more than one jobid. This counter is a usefull
shortcut for finalize and abort. */
int ompi_comm_num_dyncomm=0;
/*
* Initialize comm world/self/null/parent.
*/
int ompi_comm_init(void)
{
ompi_group_t *group;
size_t size;
/* Setup communicator array */
OBJ_CONSTRUCT(&ompi_mpi_communicators, opal_pointer_array_t);
if( OPAL_SUCCESS != opal_pointer_array_init(&ompi_mpi_communicators, 16,
OMPI_FORTRAN_HANDLE_MAX, 64) ) {
return OMPI_ERROR;
}
/* Setup f to c table (we can no longer use the cid as the fortran handle) */
OBJ_CONSTRUCT(&ompi_comm_f_to_c_table, opal_pointer_array_t);
if( OPAL_SUCCESS != opal_pointer_array_init(&ompi_comm_f_to_c_table, 8,
OMPI_FORTRAN_HANDLE_MAX, 32) ) {
return OMPI_ERROR;
}
/* Setup MPI_COMM_WORLD */
OBJ_CONSTRUCT(&ompi_mpi_comm_world, ompi_communicator_t);
assert(ompi_mpi_comm_world.comm.c_f_to_c_index == 0);
group = OBJ_NEW(ompi_group_t);
size = ompi_process_info.num_procs;
group->grp_proc_pointers = (ompi_proc_t **) calloc (size, sizeof (ompi_proc_t *));
group->grp_proc_count = size;
for (size_t i = 0 ; i < size ; ++i) {
opal_process_name_t name = {.vpid = i, .jobid = OMPI_PROC_MY_NAME->jobid};
/* look for existing ompi_proc_t that matches this name */
group->grp_proc_pointers[i] = (ompi_proc_t *) ompi_proc_lookup (name);
if (NULL == group->grp_proc_pointers[i]) {
/* set sentinel value */
group->grp_proc_pointers[i] = (ompi_proc_t *) ompi_proc_name_to_sentinel (name);
} else {
OBJ_RETAIN (group->grp_proc_pointers[i]);
}
}
OMPI_GROUP_SET_INTRINSIC (group);
OMPI_GROUP_SET_DENSE (group);
ompi_set_group_rank(group, ompi_proc_local());
ompi_mpi_comm_world.comm.c_contextid = 0;
ompi_mpi_comm_world.comm.c_id_start_index = 4;
ompi_mpi_comm_world.comm.c_id_available = 4;
ompi_mpi_comm_world.comm.c_my_rank = group->grp_my_rank;
ompi_mpi_comm_world.comm.c_local_group = group;
ompi_mpi_comm_world.comm.c_remote_group = group;
OBJ_RETAIN(ompi_mpi_comm_world.comm.c_remote_group);
ompi_mpi_comm_world.comm.c_cube_dim = opal_cube_dim((int)size);
ompi_mpi_comm_world.comm.error_handler = &ompi_mpi_errors_are_fatal.eh;
OBJ_RETAIN( &ompi_mpi_errors_are_fatal.eh );
OMPI_COMM_SET_PML_ADDED(&ompi_mpi_comm_world.comm);
opal_pointer_array_set_item (&ompi_mpi_communicators, 0, &ompi_mpi_comm_world);
MEMCHECKER (memset (ompi_mpi_comm_world.comm.c_name, 0, MPI_MAX_OBJECT_NAME));
strncpy (ompi_mpi_comm_world.comm.c_name, "MPI_COMM_WORLD",
strlen("MPI_COMM_WORLD")+1 );
ompi_mpi_comm_world.comm.c_flags |= OMPI_COMM_NAMEISSET;
ompi_mpi_comm_world.comm.c_flags |= OMPI_COMM_INTRINSIC;
/* We have to create a hash (although it is legal to leave this
filed NULL -- the attribute accessor functions will intepret
this as "there are no attributes cached on this object")
because MPI_COMM_WORLD has some predefined attributes. */
ompi_attr_hash_init(&ompi_mpi_comm_world.comm.c_keyhash);
/* Check for the binding policy used. We are only interested in
whether mapby-node has been set right now (could be extended later)
and only on MPI_COMM_WORLD, since for all other sub-communicators
it is virtually impossible to identify their layout across nodes
in the most generic sense. This is used by OMPIO for deciding which
ranks to use for aggregators
*/
opal_process_name_t wildcard = {OMPI_PROC_MY_NAME->jobid, OPAL_VPID_WILDCARD};
char *str=NULL;
int rc;
OPAL_MODEX_RECV_VALUE_OPTIONAL(rc, OPAL_PMIX_MAPBY, &wildcard, &str, OPAL_STRING);
if ( 0 == rc && NULL != str) {
if ( strstr ( str, "BYNODE") ) {
OMPI_COMM_SET_MAPBY_NODE(&ompi_mpi_comm_world.comm);
}
if (NULL != str) {
free(str);
}
}
/* Setup MPI_COMM_SELF */
OBJ_CONSTRUCT(&ompi_mpi_comm_self, ompi_communicator_t);
assert(ompi_mpi_comm_self.comm.c_f_to_c_index == 1);
group = OBJ_NEW(ompi_group_t);
group->grp_proc_pointers = ompi_proc_self(&size);
group->grp_my_rank = 0;
group->grp_proc_count = (int)size;
OMPI_GROUP_SET_INTRINSIC (group);
OMPI_GROUP_SET_DENSE (group);
ompi_mpi_comm_self.comm.c_contextid = 1;
ompi_mpi_comm_self.comm.c_id_start_index = 20;
ompi_mpi_comm_self.comm.c_id_available = 20;
ompi_mpi_comm_self.comm.c_my_rank = group->grp_my_rank;
ompi_mpi_comm_self.comm.c_local_group = group;
ompi_mpi_comm_self.comm.c_remote_group = group;
OBJ_RETAIN(ompi_mpi_comm_self.comm.c_remote_group);
ompi_mpi_comm_self.comm.error_handler = &ompi_mpi_errors_are_fatal.eh;
OBJ_RETAIN( &ompi_mpi_errors_are_fatal.eh );
OMPI_COMM_SET_PML_ADDED(&ompi_mpi_comm_self.comm);
opal_pointer_array_set_item (&ompi_mpi_communicators, 1, &ompi_mpi_comm_self);
MEMCHECKER (memset (ompi_mpi_comm_self.comm.c_name, 0, MPI_MAX_OBJECT_NAME));
strncpy(ompi_mpi_comm_self.comm.c_name,"MPI_COMM_SELF",strlen("MPI_COMM_SELF")+1);
ompi_mpi_comm_self.comm.c_flags |= OMPI_COMM_NAMEISSET;
ompi_mpi_comm_self.comm.c_flags |= OMPI_COMM_INTRINSIC;
/* We can set MPI_COMM_SELF's keyhash to NULL because it has no
predefined attributes. If a user defines an attribute on
MPI_COMM_SELF, the keyhash will automatically be created. */
ompi_mpi_comm_self.comm.c_keyhash = NULL;
/* Setup MPI_COMM_NULL */
OBJ_CONSTRUCT(&ompi_mpi_comm_null, ompi_communicator_t);
assert(ompi_mpi_comm_null.comm.c_f_to_c_index == 2);
ompi_mpi_comm_null.comm.c_local_group = &ompi_mpi_group_null.group;
ompi_mpi_comm_null.comm.c_remote_group = &ompi_mpi_group_null.group;
OBJ_RETAIN(&ompi_mpi_group_null.group);
OBJ_RETAIN(&ompi_mpi_group_null.group);
ompi_mpi_comm_null.comm.c_contextid = 2;
ompi_mpi_comm_null.comm.c_my_rank = MPI_PROC_NULL;
ompi_mpi_comm_null.comm.error_handler = &ompi_mpi_errors_are_fatal.eh;
OBJ_RETAIN( &ompi_mpi_errors_are_fatal.eh );
opal_pointer_array_set_item (&ompi_mpi_communicators, 2, &ompi_mpi_comm_null);
MEMCHECKER (memset (ompi_mpi_comm_null.comm.c_name, 0, MPI_MAX_OBJECT_NAME));
strncpy(ompi_mpi_comm_null.comm.c_name,"MPI_COMM_NULL",strlen("MPI_COMM_NULL")+1);
ompi_mpi_comm_null.comm.c_flags |= OMPI_COMM_NAMEISSET;
ompi_mpi_comm_null.comm.c_flags |= OMPI_COMM_INTRINSIC;
/* Initialize the parent communicator to MPI_COMM_NULL */
ompi_mpi_comm_parent = &ompi_mpi_comm_null.comm;
OBJ_RETAIN(&ompi_mpi_comm_null);
OBJ_RETAIN(&ompi_mpi_group_null.group);
OBJ_RETAIN(&ompi_mpi_errors_are_fatal.eh);
/* initialize communicator requests (for ompi_comm_idup) */
ompi_comm_request_init ();
return OMPI_SUCCESS;
}
ompi_communicator_t *ompi_comm_allocate ( int local_size, int remote_size )
{
ompi_communicator_t *new_comm;
/* create new communicator element */
new_comm = OBJ_NEW(ompi_communicator_t);
new_comm->super.s_info = NULL;
new_comm->c_local_group = ompi_group_allocate ( local_size );
if ( 0 < remote_size ) {
new_comm->c_remote_group = ompi_group_allocate (remote_size);
new_comm->c_flags |= OMPI_COMM_INTER;
} else {
/*
* simplifies some operations (e.g. p2p), if
* we can always use the remote group
*/
new_comm->c_remote_group = new_comm->c_local_group;
OBJ_RETAIN(new_comm->c_remote_group);
}
/* fill in the inscribing hyper-cube dimensions */
new_comm->c_cube_dim = opal_cube_dim(local_size);
return new_comm;
}
int ompi_comm_finalize(void)
{
int max, i;
ompi_communicator_t *comm;
/* Shut down MPI_COMM_SELF */
OBJ_DESTRUCT( &ompi_mpi_comm_self );
/* disconnect all dynamic communicators */
ompi_dpm_dyn_finalize();
/* Free the attributes on comm world. This is not done in the
* destructor as we delete attributes in ompi_comm_free (which
* is not called for comm world) */
if (NULL != ompi_mpi_comm_world.comm.c_keyhash) {
/* Ignore errors when deleting attributes on comm_world */
(void) ompi_attr_delete_all(COMM_ATTR, &ompi_mpi_comm_world.comm, ompi_mpi_comm_world.comm.c_keyhash);
OBJ_RELEASE(ompi_mpi_comm_world.comm.c_keyhash);
}
/* Shut down MPI_COMM_WORLD */
OBJ_DESTRUCT( &ompi_mpi_comm_world );
/* Shut down the parent communicator, if it exists */
if( ompi_mpi_comm_parent != &ompi_mpi_comm_null.comm ) {
/* Note that we pass ompi_mpi_comm_parent here
(vs. &ompi_mpi_comm_parent) because it is of type
(ompi_communicator_t*), *NOT* (ompi_communicator_t). This
is because a parent communicator is created dynamically
during init, and we just set this pointer to it. Hence, we
just pass in the pointer here. */
OBJ_DESTRUCT (ompi_mpi_comm_parent);
/* Please note, that the we did increase the reference count
for ompi_mpi_comm_null, ompi_mpi_group_null, and
ompi_mpi_errors_are_fatal in ompi_comm_init because of
ompi_mpi_comm_parent. In case a
parent communicator is really created, the ref. counters
for these objects are decreased again by one. However, in a
static scenario, we should ideally decrease the ref. counter
for these objects by one here. The problem just is, that
if the app had a parent_comm, and this has been freed/disconnected,
ompi_comm_parent points again to ompi_comm_null, the reference count
for these objects has not been increased again.
So the point is, if ompi_mpi_comm_parent == &ompi_mpi_comm_null
we do not know whether we have to decrease the ref count for
those three objects or not. Since this is a constant, non-increasing
amount of memory, we stick with the current solution for now,
namely don't do anything.
*/
}
/* Shut down MPI_COMM_NULL */
OBJ_DESTRUCT( &ompi_mpi_comm_null );
/* Check whether we have some communicators left */
max = opal_pointer_array_get_size(&ompi_mpi_communicators);
for ( i=3; i<max; i++ ) {
comm = (ompi_communicator_t *)opal_pointer_array_get_item(&ompi_mpi_communicators, i);
if ( NULL != comm ) {
/* Communicator has not been freed before finalize */
OBJ_RELEASE(comm);
comm=(ompi_communicator_t *)opal_pointer_array_get_item(&ompi_mpi_communicators, i);
if ( NULL != comm ) {
/* Still here ? */
if ( !OMPI_COMM_IS_EXTRA_RETAIN(comm)) {
/* For communicator that have been marked as "extra retain", we do not further
* enforce to decrease the reference counter once more. These "extra retain"
* communicators created e.g. by the hierarch or inter module did increase
* the reference count by one more than other communicators, on order to
* allow for deallocation with the parent communicator. Note, that
* this only occurs if the cid of the local_comm is lower than of its
* parent communicator. Read the comment in comm_activate for
* a full explanation.
*/
if ( ompi_debug_show_handle_leaks && !(OMPI_COMM_IS_FREED(comm)) ){
opal_output(0,"WARNING: MPI_Comm still allocated in MPI_Finalize\n");
ompi_comm_dump ( comm);
OBJ_RELEASE(comm);
}
}
}
}
}
OBJ_DESTRUCT (&ompi_mpi_communicators);
OBJ_DESTRUCT (&ompi_comm_f_to_c_table);
/* finalize communicator requests */
ompi_comm_request_fini ();
return OMPI_SUCCESS;
}
/********************************************************************************/
/********************************************************************************/
/********************************************************************************/
/* static functions */
static void ompi_comm_construct(ompi_communicator_t* comm)
{
comm->c_f_to_c_index = opal_pointer_array_add(&ompi_comm_f_to_c_table, comm);
comm->c_name[0] = '\0';
comm->c_contextid = MPI_UNDEFINED;
comm->c_id_available = MPI_UNDEFINED;
comm->c_id_start_index = MPI_UNDEFINED;
comm->c_flags = 0;
comm->c_my_rank = 0;
comm->c_cube_dim = 0;
comm->c_local_group = NULL;
comm->c_remote_group = NULL;
comm->error_handler = NULL;
comm->c_pml_comm = NULL;
comm->c_topo = NULL;
comm->c_coll = NULL;
comm->c_nbc_tag = MCA_COLL_BASE_TAG_NONBLOCKING_BASE;
/* A keyhash will be created if/when an attribute is cached on
this communicator */
comm->c_keyhash = NULL;
comm->errhandler_type = OMPI_ERRHANDLER_TYPE_COMM;
#ifdef OMPI_WANT_PERUSE
comm->c_peruse_handles = NULL;
#endif
OBJ_CONSTRUCT(&comm->c_lock, opal_mutex_t);
}
static void ompi_comm_destruct(ompi_communicator_t* comm)
{
/* Note that the attributes were already released on this
communicator in ompi_comm_free() (i.e., from MPI_COMM_FREE /
MPI_COMM_DISCONNECT). See the lengthy comment in
communicator/comm.c in ompi_comm_free() for the reasons why. */
/* Release the collective module */
if ( NULL != comm->c_coll ) {
mca_coll_base_comm_unselect(comm);
}
/* Tell the PML that this communicator is done.
MCA_PML_CALL(add_comm()) was called explicitly in
ompi_comm_init() when setting up COMM_WORLD and COMM_SELF; it's
called in ompi_comm_set() for all others. This means that all
communicators must be destroyed before the PML shuts down.
Also -- do not invoke the pml_del_comm if the corresponding
pml_add_comm was never invoked. This can happen in an error
situation where, for example, attributes do not copy properly
from one communicator to another and we end up destroying the
new communication while propagating the error up the stack. We
want to make it all the way up the stack to invoke the MPI
exception, not cause a seg fault in pml_del_comm because it was
never pml_add_com'ed. */
if ( MPI_COMM_NULL != comm && OMPI_COMM_IS_PML_ADDED(comm) ) {
MCA_PML_CALL(del_comm (comm));
}
/* Release topology module */
if (NULL != comm->c_topo) {
OBJ_RELEASE(comm->c_topo);
comm->c_topo = NULL;
}
if (NULL != comm->c_local_group) {
OBJ_RELEASE ( comm->c_local_group );
comm->c_local_group = NULL;
if ( OMPI_COMM_IS_INTRA(comm) ) {
/* We have to decrement the ref count on the remote group
even if it is identical to the local one in case of
intra-comm */
OBJ_RELEASE ( comm->c_remote_group );
comm->c_remote_group = NULL;
}
}
if (NULL != comm->c_remote_group) {
OBJ_RELEASE ( comm->c_remote_group );
comm->c_remote_group = NULL;
}
if (NULL != comm->error_handler) {
OBJ_RELEASE ( comm->error_handler );
comm->error_handler = NULL;
}
/* mark this cid as available */
if ( MPI_UNDEFINED != (int)comm->c_contextid &&
NULL != opal_pointer_array_get_item(&ompi_mpi_communicators,
comm->c_contextid)) {
opal_pointer_array_set_item ( &ompi_mpi_communicators,
comm->c_contextid, NULL);
}
/* reset the ompi_comm_f_to_c_table entry */
if ( MPI_UNDEFINED != comm->c_f_to_c_index &&
NULL != opal_pointer_array_get_item(&ompi_comm_f_to_c_table,
comm->c_f_to_c_index)) {
opal_pointer_array_set_item ( &ompi_comm_f_to_c_table,
comm->c_f_to_c_index, NULL);
}
OBJ_DESTRUCT(&comm->c_lock);
}
#define OMPI_COMM_SET_INFO_FN(name, flag) \
static char *ompi_comm_set_ ## name (opal_infosubscriber_t *obj, char *key, char *value) \
{ \
ompi_communicator_t *comm = (ompi_communicator_t *) obj; \
\
if (opal_str_to_bool(value)) { \
comm->c_assertions |= flag; \
} else { \
comm->c_assertions &= ~flag; \
} \
\
return OMPI_COMM_CHECK_ASSERT(comm, flag) ? "true" : "false"; \
}
OMPI_COMM_SET_INFO_FN(no_any_source, OMPI_COMM_ASSERT_NO_ANY_SOURCE)
OMPI_COMM_SET_INFO_FN(no_any_tag, OMPI_COMM_ASSERT_NO_ANY_TAG)
OMPI_COMM_SET_INFO_FN(allow_overtake, OMPI_COMM_ASSERT_ALLOW_OVERTAKE)
OMPI_COMM_SET_INFO_FN(exact_length, OMPI_COMM_ASSERT_EXACT_LENGTH)
void ompi_comm_assert_subscribe (ompi_communicator_t *comm, int32_t assert_flag)
{
switch (assert_flag) {
case OMPI_COMM_ASSERT_NO_ANY_SOURCE:
opal_infosubscribe_subscribe (&comm->super, "mpi_assert_no_any_source", "false", ompi_comm_set_no_any_source);
break;
case OMPI_COMM_ASSERT_NO_ANY_TAG:
opal_infosubscribe_subscribe (&comm->super, "mpi_assert_no_any_tag", "false", ompi_comm_set_no_any_tag);
break;
case OMPI_COMM_ASSERT_ALLOW_OVERTAKE:
opal_infosubscribe_subscribe (&comm->super, "mpi_assert_allow_overtaking", "false", ompi_comm_set_allow_overtake);
break;
case OMPI_COMM_ASSERT_EXACT_LENGTH:
opal_infosubscribe_subscribe (&comm->super, "mpi_assert_exact_length", "false", ompi_comm_set_exact_length);
break;
}
}
Показать git blame Копировать постоянную ссылку