1
1
openmpi/ompi/mca/coll/hierarch/coll_hierarch.c

750 строки
23 KiB
C
Исходник Обычный вид История

/* -*- Mode: C; c-basic-offset:4 ; -*- */
/*
* Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2007 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) 2007-2008 University of Houston. All rights reserved.
* Copyright (c) 2008 Sun Microsystems, Inc. All rights reserved.
Per RFC, bring in the following changes: * Remove paffinity, maffinity, and carto frameworks -- they've been wholly replaced by hwloc. * Move ompi_mpi_init() affinity-setting/checking code down to ORTE. * Update sm, smcuda, wv, and openib components to no longer use carto. Instead, use hwloc data. There are still optimizations possible in the sm/smcuda BTLs (i.e., making multiple mpools). Also, the old carto-based code found out how many NUMA nodes were ''available'' -- not how many were used ''in this job''. The new hwloc-using code computes the same value -- it was not updated to calculate how many NUMA nodes are used ''by this job.'' * Note that I cannot compile the smcuda and wv BTLs -- I ''think'' they're right, but they need to be verified by their owners. * The openib component now does a bunch of stuff to figure out where "near" OpenFabrics devices are. '''THIS IS A CHANGE IN DEFAULT BEHAVIOR!!''' and still needs to be verified by OpenFabrics vendors (I do not have a NUMA machine with an OpenFabrics device that is a non-uniform distance from multiple different NUMA nodes). * Completely rewrite the OMPI_Affinity_str() routine from the "affinity" mpiext extension. This extension now understands hyperthreads; the output format of it has changed a bit to reflect this new information. * Bunches of minor changes around the code base to update names/types from maffinity/paffinity-based names to hwloc-based names. * Add some helper functions into the hwloc base, mainly having to do with the fact that we have the hwloc data reporting ''all'' topology information, but sometimes you really only want the (online | available) data. This commit was SVN r26391.
2012-05-07 18:52:54 +04:00
* Copyright (c) 2012 Cisco Systems, Inc. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "coll_hierarch.h"
#include <stdio.h>
Per RFC, bring in the following changes: * Remove paffinity, maffinity, and carto frameworks -- they've been wholly replaced by hwloc. * Move ompi_mpi_init() affinity-setting/checking code down to ORTE. * Update sm, smcuda, wv, and openib components to no longer use carto. Instead, use hwloc data. There are still optimizations possible in the sm/smcuda BTLs (i.e., making multiple mpools). Also, the old carto-based code found out how many NUMA nodes were ''available'' -- not how many were used ''in this job''. The new hwloc-using code computes the same value -- it was not updated to calculate how many NUMA nodes are used ''by this job.'' * Note that I cannot compile the smcuda and wv BTLs -- I ''think'' they're right, but they need to be verified by their owners. * The openib component now does a bunch of stuff to figure out where "near" OpenFabrics devices are. '''THIS IS A CHANGE IN DEFAULT BEHAVIOR!!''' and still needs to be verified by OpenFabrics vendors (I do not have a NUMA machine with an OpenFabrics device that is a non-uniform distance from multiple different NUMA nodes). * Completely rewrite the OMPI_Affinity_str() routine from the "affinity" mpiext extension. This extension now understands hyperthreads; the output format of it has changed a bit to reflect this new information. * Bunches of minor changes around the code base to update names/types from maffinity/paffinity-based names to hwloc-based names. * Add some helper functions into the hwloc base, mainly having to do with the fact that we have the hwloc data reporting ''all'' topology information, but sometimes you really only want the (online | available) data. This commit was SVN r26391.
2012-05-07 18:52:54 +04:00
#include "opal/mca/hwloc/base/base.h"
#include "mpi.h"
#include "ompi/communicator/communicator.h"
#include "ompi/group/group.h"
#include "ompi/proc/proc.h"
#include "ompi/mca/coll/coll.h"
#include "ompi/mca/coll/base/base.h"
#include "ompi/mca/coll/base/coll_tags.h"
#include "opal/class/opal_bitmap.h"
#include "ompi/mca/bml/bml.h"
#include "ompi/mca/bml/base/base.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/mca/pml/base/base.h"
#include "ompi/mca/btl/btl.h"
/* Local functions and data */
#define HIER_MAXPROTOCOL 6
#define HIER_MAX_PROTNAMELEN 7
static int mca_coll_hierarch_max_protocol=HIER_MAXPROTOCOL;
/* Commments: need to add ofud, portals and sctp into this list! */
static char hier_prot[HIER_MAXPROTOCOL][HIER_MAX_PROTNAMELEN]={"0","tcp","udapl","mx","openib","sm"};
static void mca_coll_hierarch_checkfor_component (struct ompi_communicator_t *comm,
int component_level,
char *component_name,
int *key, int *ncount);
static void mca_coll_hierarch_checkfor_sm (struct ompi_communicator_t *comm,
int *color,
int *ncount);
static void mca_coll_hierarch_dump_struct ( mca_coll_hierarch_module_t *c);
/*
* Initial query function that is invoked during MPI_INIT, allowing
* this module to indicate what level of thread support it provides.
*/
int mca_coll_hierarch_init_query(bool allow_hierarch_user_threads,
bool have_hidden_user_threads)
{
/* Don't ask. All done */
return OMPI_SUCCESS;
}
/*
* Invoked when there's a new communicator that has been created.
* Look at the communicator and decide which set of functions and
* priority we want to return.
*/
mca_coll_base_module_t *
mca_coll_hierarch_comm_query(struct ompi_communicator_t *comm, int *priority )
{
int size, rank;
int color, ncount=0, maxncount;
int level;
int ret=OMPI_SUCCESS;
int ignore_sm=0;
int detection_alg=0;
mca_coll_hierarch_module_t *hierarch_module;
/* This module only works for intra-communicators at the moment */
if (OMPI_COMM_IS_INTER(comm)) {
return NULL;
}
/* Get the priority level attached to this module. If priority is less
* than or equal to 0, then the module is unavailable. */
*priority = mca_coll_hierarch_priority_param;
if (0 >= mca_coll_hierarch_priority_param) {
return NULL;
}
/* This module only works with ob1 pml respectively btl's. Opt out
of we use cm/mtl's. */
if ( strcmp ( mca_pml_base_selected_component.pmlm_version.mca_component_name, "ob1")) {
return NULL;
}
size = ompi_comm_size(comm);
if (size < 3) {
/* No need for hierarchical collectives for 1 or 2 procs. */
return NULL;
}
hierarch_module = OBJ_NEW(mca_coll_hierarch_module_t);
if (NULL == hierarch_module) {
return NULL;
}
hierarch_module->super.coll_module_enable = mca_coll_hierarch_module_enable;
hierarch_module->super.ft_event = mca_coll_hierarch_ft_event;
hierarch_module->super.coll_allgather = NULL;
hierarch_module->super.coll_allgatherv = NULL;
hierarch_module->super.coll_allreduce = mca_coll_hierarch_allreduce_intra;
hierarch_module->super.coll_alltoall = NULL;
hierarch_module->super.coll_alltoallv = NULL;
hierarch_module->super.coll_alltoallw = NULL;
hierarch_module->super.coll_barrier = mca_coll_hierarch_barrier_intra;
hierarch_module->super.coll_bcast = mca_coll_hierarch_bcast_intra;
hierarch_module->super.coll_exscan = NULL;
hierarch_module->super.coll_gather = NULL;
hierarch_module->super.coll_gatherv = NULL;
hierarch_module->super.coll_reduce = mca_coll_hierarch_reduce_intra;
hierarch_module->super.coll_reduce_scatter = NULL;
hierarch_module->super.coll_scan = NULL;
hierarch_module->super.coll_scatter = NULL;
hierarch_module->super.coll_scatterv = NULL;
/* Check whether we should ignore sm. This might be necessary to take advantage
of the some ib or gm collectives. */
ignore_sm = mca_coll_hierarch_ignore_sm_param;
rank = ompi_comm_rank(comm);
hierarch_module->hier_num_colorarr = size;
hierarch_module->hier_colorarr = (int *) malloc ( sizeof(int) * size);
if ( NULL == hierarch_module->hier_colorarr ) {
*priority = 0;
return NULL;
}
/*
* walk through the list of registered protocols, and check which one
* is feasible.
* Later we start with level=0, and introduce the multi-cell check
*/
if ( ignore_sm ) {
mca_coll_hierarch_max_protocol = HIER_MAXPROTOCOL - 1;
}
/* if number of levels is not specified, or if it is specified as ALL_LEVELS,
* proceed in the usual way
*/
detection_alg = mca_coll_hierarch_detection_alg_param;
if( TWO_LEVELS == detection_alg ) {
mca_coll_hierarch_max_protocol = 2;
if ( mca_coll_hierarch_verbose_param ) {
printf("Using two level hierarchy detection\n");
}
}
for ( level = mca_coll_hierarch_max_protocol - 1; level >0 ; level--) {
if ( ALL_LEVELS == detection_alg ) {
mca_coll_hierarch_checkfor_component ( comm,
level,
hier_prot[level],
&color,
&ncount);
}
else if (TWO_LEVELS == detection_alg ) {
mca_coll_hierarch_checkfor_sm ( comm, &color, &ncount );
}
/* This is probably a no-no! but for the moment we agreed with Jeff,
** that this might be the best solution. These functions emulate an
** allreduce and an allgather.
*/
ret = mca_coll_hierarch_allreduce_tmp (&ncount, &maxncount, 1, MPI_INT,
MPI_MAX, comm );
if ( OMPI_SUCCESS != ret ) {
return NULL;
}
if ( 0 == maxncount ) {
if ( mca_coll_hierarch_verbose_param ) {
printf("%s:%d: nobody talks with %s. Continuing to next level.\n",
comm->c_name, rank, hier_prot[level]);
}
continue;
}
else if ( maxncount == (size-1) ) {
/*
* everybody can talk to every other process with this protocol,
* no need to continue in the hierarchy tree and for the
* hierarchical component.
* Its (size-1) because we do not count ourselves.
* maxncount[1] should be zero.
*/
if ( mca_coll_hierarch_verbose_param ) {
if ( ALL_LEVELS == detection_alg ) {
printf("%s:%d: everybody talks with %s. No need to continue\n",
comm->c_name, rank, hier_prot[level]);
}
else if ( TWO_LEVELS == detection_alg ) {
printf("%s:%d: everybody talks with sm. No need to continue\n",
comm->c_name, rank );
}
}
goto exit;
}
else {
if ( mca_coll_hierarch_verbose_param ) {
printf("%s:%d: %d procs talk with %s. Use this protocol, key %d\n",
comm->c_name, rank, maxncount, hier_prot[level], color);
}
ret = mca_coll_hierarch_allgather_tmp (&color, 1, MPI_INT,
hierarch_module->hier_colorarr, 1,
MPI_INT, comm );
if ( OMPI_SUCCESS != ret ) {
return NULL;
}
hierarch_module->hier_level = level;
return &(hierarch_module->super);
}
}
exit:
*priority = 0;
return NULL;
}
/*
* Init module on the communicator
*/
int mca_coll_hierarch_module_enable (mca_coll_base_module_t *module,
struct ompi_communicator_t *comm)
{
int color;
int size, rank, ret=OMPI_SUCCESS;
struct ompi_communicator_t *lcomm=NULL;
struct ompi_communicator_t *llcomm=NULL;
struct mca_coll_hierarch_llead_t *llead=NULL;
mca_coll_hierarch_module_t *hierarch_module = (mca_coll_hierarch_module_t *) module;
rank = ompi_comm_rank(comm);
size = ompi_comm_size(comm);
color = hierarch_module->hier_colorarr[rank];
/* Generate the subcommunicator based on the color returned by
the previous function. */
ret = ompi_comm_split ( comm, color, rank, &lcomm, 0 );
if ( OMPI_SUCCESS != ret ) {
goto exit;
}
This commit fixes a bug on how to deal with the potential if a 'dependent' communicator that we created has a lower CID than the parent comm. This can happen when using the hierarch collective communication module or for inter-communicators (since we make a duplicate of the original communicator). This is not a problem as long as the user calls MPI_Comm_free on the parent communicator. However, if the communicators are not freed by the user but released by Open MPI in MPI_Finalize, we walk through the list of still available communicators and free them one by one. Thus, local_comm is freed before the actual inter-communicator. However, the local_comm pointer in the inter communicator will still contain the 'previous' address of the local_comm and thus this will lead to a segmentation violation. In order to prevent that from happening, we increase the reference counter local_comm by one if its CID is lower than the parent. We cannot increase however its reference counter if the CID of local_comm is larger than the CID of the inter communicators, since a regular MPI_Comm_free would leave in that the case the local_comm hanging around and thus we would not recycle CID's properly, which was the reason and the cause for this trouble. This commit fixes tickets 2094 and 2166. Note however, that I want to close them manually, since a slightly different patch is required for the 1.4 series. This commit will have to be applied for the 1.5 series. And I will need a volunteer to review it. This commit was SVN r22671.
2010-02-20 02:45:30 +03:00
if ( OMPI_COMM_CID_IS_LOWER ( lcomm, comm ) ) {
/* Mark the communicator as 'extra retain' and increase the
This commit fixes a bug on how to deal with the potential if a 'dependent' communicator that we created has a lower CID than the parent comm. This can happen when using the hierarch collective communication module or for inter-communicators (since we make a duplicate of the original communicator). This is not a problem as long as the user calls MPI_Comm_free on the parent communicator. However, if the communicators are not freed by the user but released by Open MPI in MPI_Finalize, we walk through the list of still available communicators and free them one by one. Thus, local_comm is freed before the actual inter-communicator. However, the local_comm pointer in the inter communicator will still contain the 'previous' address of the local_comm and thus this will lead to a segmentation violation. In order to prevent that from happening, we increase the reference counter local_comm by one if its CID is lower than the parent. We cannot increase however its reference counter if the CID of local_comm is larger than the CID of the inter communicators, since a regular MPI_Comm_free would leave in that the case the local_comm hanging around and thus we would not recycle CID's properly, which was the reason and the cause for this trouble. This commit fixes tickets 2094 and 2166. Note however, that I want to close them manually, since a slightly different patch is required for the 1.4 series. This commit will have to be applied for the 1.5 series. And I will need a volunteer to review it. This commit was SVN r22671.
2010-02-20 02:45:30 +03:00
reference count by one more. See ompi_comm_activate
for detailed comments
*/
OMPI_COMM_SET_EXTRA_RETAIN (lcomm);
This commit fixes a bug on how to deal with the potential if a 'dependent' communicator that we created has a lower CID than the parent comm. This can happen when using the hierarch collective communication module or for inter-communicators (since we make a duplicate of the original communicator). This is not a problem as long as the user calls MPI_Comm_free on the parent communicator. However, if the communicators are not freed by the user but released by Open MPI in MPI_Finalize, we walk through the list of still available communicators and free them one by one. Thus, local_comm is freed before the actual inter-communicator. However, the local_comm pointer in the inter communicator will still contain the 'previous' address of the local_comm and thus this will lead to a segmentation violation. In order to prevent that from happening, we increase the reference counter local_comm by one if its CID is lower than the parent. We cannot increase however its reference counter if the CID of local_comm is larger than the CID of the inter communicators, since a regular MPI_Comm_free would leave in that the case the local_comm hanging around and thus we would not recycle CID's properly, which was the reason and the cause for this trouble. This commit fixes tickets 2094 and 2166. Note however, that I want to close them manually, since a slightly different patch is required for the 1.4 series. This commit will have to be applied for the 1.5 series. And I will need a volunteer to review it. This commit was SVN r22671.
2010-02-20 02:45:30 +03:00
OBJ_RETAIN(lcomm);
}
hierarch_module->hier_comm = comm;
hierarch_module->hier_lcomm = lcomm;
hierarch_module->hier_num_reqs = 2 * size;
hierarch_module->hier_reqs = (ompi_request_t **) malloc (sizeof(ompi_request_t)*size*2);
if ( NULL == hierarch_module->hier_reqs ) {
goto exit;
}
/* allocate a certain number of the hierarch_llead structures, which store
information about local leader and the according subcommunicators
*/
llead = (struct mca_coll_hierarch_llead_t * ) malloc (
sizeof(struct mca_coll_hierarch_llead_t));
if ( NULL == llead ) {
goto exit;
}
/* These two routines set all relevant entries in the mca_coll_base_comm_t
* structure. The first one makes all entries which are independent of the
* offset (and have to be done only once per module. The second one is
* depending on the offset, and has to be called therefore every time we need
* a new llcomm
*/
mca_coll_hierarch_get_llr ( hierarch_module );
mca_coll_hierarch_get_all_lleaders ( rank, hierarch_module, llead, 1 );
/* Generate the lleader communicator assuming that all lleaders are the first
process in the list of processes with the same color. A function generating
other lleader-comms will follow soon. */
This commit fixes a bug on how to deal with the potential if a 'dependent' communicator that we created has a lower CID than the parent comm. This can happen when using the hierarch collective communication module or for inter-communicators (since we make a duplicate of the original communicator). This is not a problem as long as the user calls MPI_Comm_free on the parent communicator. However, if the communicators are not freed by the user but released by Open MPI in MPI_Finalize, we walk through the list of still available communicators and free them one by one. Thus, local_comm is freed before the actual inter-communicator. However, the local_comm pointer in the inter communicator will still contain the 'previous' address of the local_comm and thus this will lead to a segmentation violation. In order to prevent that from happening, we increase the reference counter local_comm by one if its CID is lower than the parent. We cannot increase however its reference counter if the CID of local_comm is larger than the CID of the inter communicators, since a regular MPI_Comm_free would leave in that the case the local_comm hanging around and thus we would not recycle CID's properly, which was the reason and the cause for this trouble. This commit fixes tickets 2094 and 2166. Note however, that I want to close them manually, since a slightly different patch is required for the 1.4 series. This commit will have to be applied for the 1.5 series. And I will need a volunteer to review it. This commit was SVN r22671.
2010-02-20 02:45:30 +03:00
color = MPI_UNDEFINED;
if ( llead->am_lleader ) {
color = 1;
}
ret = ompi_comm_split ( comm, color, rank, &llcomm, 0);
if ( OMPI_SUCCESS != ret ) {
goto exit;
}
This commit fixes a bug on how to deal with the potential if a 'dependent' communicator that we created has a lower CID than the parent comm. This can happen when using the hierarch collective communication module or for inter-communicators (since we make a duplicate of the original communicator). This is not a problem as long as the user calls MPI_Comm_free on the parent communicator. However, if the communicators are not freed by the user but released by Open MPI in MPI_Finalize, we walk through the list of still available communicators and free them one by one. Thus, local_comm is freed before the actual inter-communicator. However, the local_comm pointer in the inter communicator will still contain the 'previous' address of the local_comm and thus this will lead to a segmentation violation. In order to prevent that from happening, we increase the reference counter local_comm by one if its CID is lower than the parent. We cannot increase however its reference counter if the CID of local_comm is larger than the CID of the inter communicators, since a regular MPI_Comm_free would leave in that the case the local_comm hanging around and thus we would not recycle CID's properly, which was the reason and the cause for this trouble. This commit fixes tickets 2094 and 2166. Note however, that I want to close them manually, since a slightly different patch is required for the 1.4 series. This commit will have to be applied for the 1.5 series. And I will need a volunteer to review it. This commit was SVN r22671.
2010-02-20 02:45:30 +03:00
if ( OMPI_COMM_CID_IS_LOWER ( llcomm, comm ) ) {
/* Mark the communicator as 'extra retain' and increase the
This commit fixes a bug on how to deal with the potential if a 'dependent' communicator that we created has a lower CID than the parent comm. This can happen when using the hierarch collective communication module or for inter-communicators (since we make a duplicate of the original communicator). This is not a problem as long as the user calls MPI_Comm_free on the parent communicator. However, if the communicators are not freed by the user but released by Open MPI in MPI_Finalize, we walk through the list of still available communicators and free them one by one. Thus, local_comm is freed before the actual inter-communicator. However, the local_comm pointer in the inter communicator will still contain the 'previous' address of the local_comm and thus this will lead to a segmentation violation. In order to prevent that from happening, we increase the reference counter local_comm by one if its CID is lower than the parent. We cannot increase however its reference counter if the CID of local_comm is larger than the CID of the inter communicators, since a regular MPI_Comm_free would leave in that the case the local_comm hanging around and thus we would not recycle CID's properly, which was the reason and the cause for this trouble. This commit fixes tickets 2094 and 2166. Note however, that I want to close them manually, since a slightly different patch is required for the 1.4 series. This commit will have to be applied for the 1.5 series. And I will need a volunteer to review it. This commit was SVN r22671.
2010-02-20 02:45:30 +03:00
reference count by one more. See ompi_comm_activate
for detailed explanation.
*/
OMPI_COMM_SET_EXTRA_RETAIN (llcomm);
This commit fixes a bug on how to deal with the potential if a 'dependent' communicator that we created has a lower CID than the parent comm. This can happen when using the hierarch collective communication module or for inter-communicators (since we make a duplicate of the original communicator). This is not a problem as long as the user calls MPI_Comm_free on the parent communicator. However, if the communicators are not freed by the user but released by Open MPI in MPI_Finalize, we walk through the list of still available communicators and free them one by one. Thus, local_comm is freed before the actual inter-communicator. However, the local_comm pointer in the inter communicator will still contain the 'previous' address of the local_comm and thus this will lead to a segmentation violation. In order to prevent that from happening, we increase the reference counter local_comm by one if its CID is lower than the parent. We cannot increase however its reference counter if the CID of local_comm is larger than the CID of the inter communicators, since a regular MPI_Comm_free would leave in that the case the local_comm hanging around and thus we would not recycle CID's properly, which was the reason and the cause for this trouble. This commit fixes tickets 2094 and 2166. Note however, that I want to close them manually, since a slightly different patch is required for the 1.4 series. This commit will have to be applied for the 1.5 series. And I will need a volunteer to review it. This commit was SVN r22671.
2010-02-20 02:45:30 +03:00
OBJ_RETAIN(llcomm);
}
llead->llcomm = llcomm;
/* Store it now on the data structure */
OBJ_CONSTRUCT(&(hierarch_module->hier_llead), opal_pointer_array_t);
opal_pointer_array_add ( &(hierarch_module->hier_llead), llead);
if ( mca_coll_hierarch_verbose_param ) {
mca_coll_hierarch_dump_struct (hierarch_module);
}
exit:
if ( OMPI_SUCCESS != ret ) {
if (NULL != llead) {
free(llead);
}
ompi_comm_free ( &lcomm );
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
int mca_coll_hierarch_get_all_lleaders ( int rank, mca_coll_hierarch_module_t *hierarch_module,
struct mca_coll_hierarch_llead_t * llead,
int offset )
{
int i, j, ret=OMPI_SUCCESS;
int *cntarr=NULL;
int mycolor;
cntarr = (int *)calloc (1, sizeof (int)* hierarch_module->hier_num_lleaders );
if ( NULL == cntarr ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
llead->lleaders = (int *) malloc (sizeof(int) * hierarch_module->hier_num_lleaders);
if ( NULL == llead->lleaders ) {
ret = OMPI_ERR_OUT_OF_RESOURCE;
goto exit;
}
llead->offset = offset;
for ( i=0; i < hierarch_module->hier_num_lleaders; i++ ) {
if ( MPI_UNDEFINED == hierarch_module->hier_llr[i] ) {
cntarr[i] = 1;
llead->lleaders[i] = MPI_UNDEFINED;
}
}
for ( i=0; i<hierarch_module->hier_num_colorarr; i++) {
if ( MPI_UNDEFINED == hierarch_module->hier_colorarr[i] ) {
continue;
}
for ( j=0; j<hierarch_module->hier_num_lleaders; j++) {
if ( cntarr[j] >= offset ) {
continue;
}
if ( hierarch_module->hier_colorarr[i] == hierarch_module->hier_llr[j]) {
cntarr[j]++;
llead->lleaders[j] = i;
break;
}
}
}
mycolor = hierarch_module->hier_colorarr[rank];
if ( MPI_UNDEFINED == mycolor ) {
llead->am_lleader = 1;
llead->my_lleader = MPI_UNDEFINED;
}
else {
This commit fixes a bug on how to deal with the potential if a 'dependent' communicator that we created has a lower CID than the parent comm. This can happen when using the hierarch collective communication module or for inter-communicators (since we make a duplicate of the original communicator). This is not a problem as long as the user calls MPI_Comm_free on the parent communicator. However, if the communicators are not freed by the user but released by Open MPI in MPI_Finalize, we walk through the list of still available communicators and free them one by one. Thus, local_comm is freed before the actual inter-communicator. However, the local_comm pointer in the inter communicator will still contain the 'previous' address of the local_comm and thus this will lead to a segmentation violation. In order to prevent that from happening, we increase the reference counter local_comm by one if its CID is lower than the parent. We cannot increase however its reference counter if the CID of local_comm is larger than the CID of the inter communicators, since a regular MPI_Comm_free would leave in that the case the local_comm hanging around and thus we would not recycle CID's properly, which was the reason and the cause for this trouble. This commit fixes tickets 2094 and 2166. Note however, that I want to close them manually, since a slightly different patch is required for the 1.4 series. This commit will have to be applied for the 1.5 series. And I will need a volunteer to review it. This commit was SVN r22671.
2010-02-20 02:45:30 +03:00
llead->am_lleader = 0;
for ( i=0; i< hierarch_module->hier_num_lleaders; i++ ) {
if ( hierarch_module->hier_llr[i] == mycolor ) {
llead->my_lleader = cntarr[i]-1;
if ( llead->lleaders[i] == rank ) {
llead->am_lleader = 1;
}
break;
}
}
}
exit:
if ( NULL != cntarr ) {
free ( cntarr );
}
return ret;
}
int mca_coll_hierarch_get_llr ( mca_coll_hierarch_module_t *hierarch_module )
{
int i, j, cnt, found;
int ncount;
ncount = mca_coll_hierarch_count_lleaders ( hierarch_module->hier_num_colorarr,
hierarch_module->hier_colorarr);
hierarch_module->hier_num_lleaders = ncount;
hierarch_module->hier_llr = (int *) malloc ( hierarch_module->hier_num_lleaders * sizeof(int));
hierarch_module->hier_max_offset = (int *) calloc ( 1, hierarch_module->hier_num_lleaders * sizeof(int));
if ( ( NULL == hierarch_module->hier_llr) || ( NULL == hierarch_module->hier_max_offset )) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
hierarch_module->hier_llr[0] = hierarch_module->hier_colorarr[0];
hierarch_module->hier_max_offset[0]=1;
for ( cnt=1, i=1; i<hierarch_module->hier_num_colorarr; i++ ) {
if ( MPI_UNDEFINED == hierarch_module->hier_colorarr[i] ) {
hierarch_module->hier_llr[cnt] = hierarch_module->hier_colorarr[i];
hierarch_module->hier_max_offset[cnt] = 1;
cnt++;
continue;
}
for ( found=0, j=0; j<cnt; j++ ) {
if ( hierarch_module->hier_llr[j] == hierarch_module->hier_colorarr[i]) {
hierarch_module->hier_max_offset[j]++;
found = 1;
break;
}
}
if ( !found ) {
hierarch_module->hier_llr[cnt] = hierarch_module->hier_colorarr[i];
hierarch_module->hier_max_offset[cnt]++;
cnt++;
}
}
return OMPI_SUCCESS;
}
struct ompi_communicator_t* mca_coll_hierarch_get_llcomm (int root,
mca_coll_hierarch_module_t *hierarch_module,
int* llroot,
int* lroot)
{
struct ompi_communicator_t *llcomm=NULL;
struct ompi_group_t *llgroup=NULL;
struct ompi_group_t *group=NULL;
struct mca_coll_hierarch_llead_t *llead=NULL;
int found, i, rc, num_llead, offset;
int rank = ompi_comm_rank (hierarch_module->hier_comm);
This commit fixes a bug on how to deal with the potential if a 'dependent' communicator that we created has a lower CID than the parent comm. This can happen when using the hierarch collective communication module or for inter-communicators (since we make a duplicate of the original communicator). This is not a problem as long as the user calls MPI_Comm_free on the parent communicator. However, if the communicators are not freed by the user but released by Open MPI in MPI_Finalize, we walk through the list of still available communicators and free them one by one. Thus, local_comm is freed before the actual inter-communicator. However, the local_comm pointer in the inter communicator will still contain the 'previous' address of the local_comm and thus this will lead to a segmentation violation. In order to prevent that from happening, we increase the reference counter local_comm by one if its CID is lower than the parent. We cannot increase however its reference counter if the CID of local_comm is larger than the CID of the inter communicators, since a regular MPI_Comm_free would leave in that the case the local_comm hanging around and thus we would not recycle CID's properly, which was the reason and the cause for this trouble. This commit fixes tickets 2094 and 2166. Note however, that I want to close them manually, since a slightly different patch is required for the 1.4 series. This commit will have to be applied for the 1.5 series. And I will need a volunteer to review it. This commit was SVN r22671.
2010-02-20 02:45:30 +03:00
int color;
/* determine what our offset of root is in the colorarr */
offset = mca_coll_hierarch_get_offset ( root,
hierarch_module->hier_num_colorarr,
hierarch_module->hier_colorarr );
num_llead = opal_pointer_array_get_size ( &(hierarch_module->hier_llead) );
for ( found=0, i=0; i < num_llead; i++ ) {
llead = (struct mca_coll_hierarch_llead_t *) opal_pointer_array_get_item (
&(hierarch_module->hier_llead), i );
if ( NULL == llead ) {
continue;
}
if (llead->offset == offset ) {
found = 1;
break;
}
#if 0
else if () {
/* the offset of root = maxoffset of this color and
* the offset on llead is larger then offset of root.
* then we can also use this llead structure
*/
}
#endif
}
if ( !found ) {
/* allocate a new llead element */
llead = (struct mca_coll_hierarch_llead_t *) malloc (
sizeof(struct mca_coll_hierarch_llead_t));
if ( NULL == llead ) {
return NULL;
}
/* generate the list of lleaders with this offset */
mca_coll_hierarch_get_all_lleaders ( rank, hierarch_module, llead, offset );
This commit fixes a bug on how to deal with the potential if a 'dependent' communicator that we created has a lower CID than the parent comm. This can happen when using the hierarch collective communication module or for inter-communicators (since we make a duplicate of the original communicator). This is not a problem as long as the user calls MPI_Comm_free on the parent communicator. However, if the communicators are not freed by the user but released by Open MPI in MPI_Finalize, we walk through the list of still available communicators and free them one by one. Thus, local_comm is freed before the actual inter-communicator. However, the local_comm pointer in the inter communicator will still contain the 'previous' address of the local_comm and thus this will lead to a segmentation violation. In order to prevent that from happening, we increase the reference counter local_comm by one if its CID is lower than the parent. We cannot increase however its reference counter if the CID of local_comm is larger than the CID of the inter communicators, since a regular MPI_Comm_free would leave in that the case the local_comm hanging around and thus we would not recycle CID's properly, which was the reason and the cause for this trouble. This commit fixes tickets 2094 and 2166. Note however, that I want to close them manually, since a slightly different patch is required for the 1.4 series. This commit will have to be applied for the 1.5 series. And I will need a volunteer to review it. This commit was SVN r22671.
2010-02-20 02:45:30 +03:00
color = MPI_UNDEFINED;
if ( llead->am_lleader ) {
color = 1;
}
/* create new lleader subcommunicator */
This commit fixes a bug on how to deal with the potential if a 'dependent' communicator that we created has a lower CID than the parent comm. This can happen when using the hierarch collective communication module or for inter-communicators (since we make a duplicate of the original communicator). This is not a problem as long as the user calls MPI_Comm_free on the parent communicator. However, if the communicators are not freed by the user but released by Open MPI in MPI_Finalize, we walk through the list of still available communicators and free them one by one. Thus, local_comm is freed before the actual inter-communicator. However, the local_comm pointer in the inter communicator will still contain the 'previous' address of the local_comm and thus this will lead to a segmentation violation. In order to prevent that from happening, we increase the reference counter local_comm by one if its CID is lower than the parent. We cannot increase however its reference counter if the CID of local_comm is larger than the CID of the inter communicators, since a regular MPI_Comm_free would leave in that the case the local_comm hanging around and thus we would not recycle CID's properly, which was the reason and the cause for this trouble. This commit fixes tickets 2094 and 2166. Note however, that I want to close them manually, since a slightly different patch is required for the 1.4 series. This commit will have to be applied for the 1.5 series. And I will need a volunteer to review it. This commit was SVN r22671.
2010-02-20 02:45:30 +03:00
rc = ompi_comm_split ( hierarch_module->hier_comm, color, root, &llcomm, 0);
if ( OMPI_SUCCESS != rc ) {
return NULL;
}
This commit fixes a bug on how to deal with the potential if a 'dependent' communicator that we created has a lower CID than the parent comm. This can happen when using the hierarch collective communication module or for inter-communicators (since we make a duplicate of the original communicator). This is not a problem as long as the user calls MPI_Comm_free on the parent communicator. However, if the communicators are not freed by the user but released by Open MPI in MPI_Finalize, we walk through the list of still available communicators and free them one by one. Thus, local_comm is freed before the actual inter-communicator. However, the local_comm pointer in the inter communicator will still contain the 'previous' address of the local_comm and thus this will lead to a segmentation violation. In order to prevent that from happening, we increase the reference counter local_comm by one if its CID is lower than the parent. We cannot increase however its reference counter if the CID of local_comm is larger than the CID of the inter communicators, since a regular MPI_Comm_free would leave in that the case the local_comm hanging around and thus we would not recycle CID's properly, which was the reason and the cause for this trouble. This commit fixes tickets 2094 and 2166. Note however, that I want to close them manually, since a slightly different patch is required for the 1.4 series. This commit will have to be applied for the 1.5 series. And I will need a volunteer to review it. This commit was SVN r22671.
2010-02-20 02:45:30 +03:00
if ( OMPI_COMM_CID_IS_LOWER ( llcomm, hierarch_module->hier_comm ) ) {
/* Mark the communicator as 'extra retain' and increase the
This commit fixes a bug on how to deal with the potential if a 'dependent' communicator that we created has a lower CID than the parent comm. This can happen when using the hierarch collective communication module or for inter-communicators (since we make a duplicate of the original communicator). This is not a problem as long as the user calls MPI_Comm_free on the parent communicator. However, if the communicators are not freed by the user but released by Open MPI in MPI_Finalize, we walk through the list of still available communicators and free them one by one. Thus, local_comm is freed before the actual inter-communicator. However, the local_comm pointer in the inter communicator will still contain the 'previous' address of the local_comm and thus this will lead to a segmentation violation. In order to prevent that from happening, we increase the reference counter local_comm by one if its CID is lower than the parent. We cannot increase however its reference counter if the CID of local_comm is larger than the CID of the inter communicators, since a regular MPI_Comm_free would leave in that the case the local_comm hanging around and thus we would not recycle CID's properly, which was the reason and the cause for this trouble. This commit fixes tickets 2094 and 2166. Note however, that I want to close them manually, since a slightly different patch is required for the 1.4 series. This commit will have to be applied for the 1.5 series. And I will need a volunteer to review it. This commit was SVN r22671.
2010-02-20 02:45:30 +03:00
reference count by one more. See ompi_comm_activate
for detailed explanation. */
OMPI_COMM_SET_EXTRA_RETAIN (llcomm);
This commit fixes a bug on how to deal with the potential if a 'dependent' communicator that we created has a lower CID than the parent comm. This can happen when using the hierarch collective communication module or for inter-communicators (since we make a duplicate of the original communicator). This is not a problem as long as the user calls MPI_Comm_free on the parent communicator. However, if the communicators are not freed by the user but released by Open MPI in MPI_Finalize, we walk through the list of still available communicators and free them one by one. Thus, local_comm is freed before the actual inter-communicator. However, the local_comm pointer in the inter communicator will still contain the 'previous' address of the local_comm and thus this will lead to a segmentation violation. In order to prevent that from happening, we increase the reference counter local_comm by one if its CID is lower than the parent. We cannot increase however its reference counter if the CID of local_comm is larger than the CID of the inter communicators, since a regular MPI_Comm_free would leave in that the case the local_comm hanging around and thus we would not recycle CID's properly, which was the reason and the cause for this trouble. This commit fixes tickets 2094 and 2166. Note however, that I want to close them manually, since a slightly different patch is required for the 1.4 series. This commit will have to be applied for the 1.5 series. And I will need a volunteer to review it. This commit was SVN r22671.
2010-02-20 02:45:30 +03:00
OBJ_RETAIN(llcomm);
}
llead->llcomm = llcomm;
/* Store the new element on the hierarch_module struct */
opal_pointer_array_add ( &(hierarch_module->hier_llead), llead);
}
llcomm = llead->llcomm;
*lroot = llead->my_lleader;
*llroot = MPI_UNDEFINED;
if ( MPI_COMM_NULL != llcomm ) {
group = hierarch_module->hier_comm->c_local_group;
llgroup = llcomm->c_local_group;
rc = ompi_group_translate_ranks ( group, 1, &root, llgroup, llroot);
if ( OMPI_SUCCESS != rc ) {
return NULL;
}
}
return llcomm;
}
/**********************************************************************/
/**********************************************************************/
/**********************************************************************/
static void
mca_coll_hierarch_checkfor_sm ( struct ompi_communicator_t *comm, int *color, int *ncount )
{
int i, size;
int lncount=0;
struct ompi_proc_t** procs=NULL;
struct ompi_proc_t* my_proc=NULL;
*color = -1;
size = ompi_comm_size(comm);
my_proc = ompi_proc_local();
procs = comm->c_local_group->grp_proc_pointers;
for ( i = 0 ; i < size ; i++) {
if ( procs[i]->proc_name.jobid == my_proc->proc_name.jobid &&
( OPAL_PROC_ON_LOCAL_NODE(procs[i]->proc_flags)) ) {
lncount++;
if ( *color == -1){
*color = i;
}
}
}
/* we need to decrease ncount in order to make the other allreduce/allgather
operations work */
lncount--;
*ncount = lncount;
return;
}
/* This function checks how many processes are using the component
'component_name' for communication and returns this count in
'ncount'. Furthermore it returns a 'key', which can be used to split
the communicator into subgroups, such that the new communicators
will definitly have all processes communicate with this component.
Oct 13: the algorithm has been modified such that it returns the
number of processes using the specified component and the number
of processes to which an even 'faster' protocol is being used. (Faster
specified in this context as being further up in the list of
hier_prot protocols specified at the beginning of this file).
*/
static void
mca_coll_hierarch_checkfor_component ( struct ompi_communicator_t *comm,
int component_level,
char *component_name,
int *key,
int *ncount )
{
opal_bitmap_t reachable;
ompi_proc_t **procs=NULL;
struct mca_bml_base_btl_array_t *bml_btl_array=NULL;
mca_bml_base_btl_t *bml_btl=NULL;
mca_btl_base_component_t *btl=NULL;
int i, size, rc;
int counter=0;
int firstproc=999999;
int rank = -1;
int use_rdma=0;
/* default values in case an error occurs */
*ncount=0;
*key=MPI_UNDEFINED;
/* Shall we check the the rdma list instead of send-list in the endpoint-structure? */
use_rdma = mca_coll_hierarch_use_rdma_param;
size = ompi_comm_size ( comm );
rank = ompi_comm_rank ( comm );
OBJ_CONSTRUCT(&reachable, opal_bitmap_t);
rc = opal_bitmap_init(&reachable, size);
if(OMPI_SUCCESS != rc) {
return;
}
procs = comm->c_local_group->grp_proc_pointers;
rc = mca_bml.bml_add_procs ( size, procs, &reachable );
if(OMPI_SUCCESS != rc) {
return;
}
for ( i=0; i<size; i++ ) {
if ( rank == i ) {
/* skip myself */
continue;
}
if ( use_rdma ) {
bml_btl_array = &(procs[i]->proc_bml->btl_rdma);
}
else {
bml_btl_array = &(procs[i]->proc_bml->btl_send);
}
bml_btl = mca_bml_base_btl_array_get_index ( bml_btl_array, 0 );
btl = bml_btl->btl->btl_component;
/* sanity check */
if ( strcmp(btl->btl_version.mca_type_name,"btl") ) {
printf("Oops, got the wrong component! type_name = %s\n",
btl->btl_version.mca_type_name );
}
/* check for the required component */
if (! strcmp (btl->btl_version.mca_component_name, component_name)){
counter++;
if (i<firstproc ) {
firstproc = i;
}
continue;
}
}
*ncount = counter;
/* final decision */
if ( counter == 0 ) {
/* this is the section indicating, that we are not
using this component */
firstproc = MPI_UNDEFINED;
}
else {
if ( rank < firstproc ) {
firstproc = rank;
}
}
*key = firstproc;
2005-10-13 00:56:29 +04:00
return;
}
/********************************************************************************/
/********************************************************************************/
/********************************************************************************/
static void mca_coll_hierarch_dump_struct ( mca_coll_hierarch_module_t *c)
{
int i, j;
int rank;
struct mca_coll_hierarch_llead_t *current=NULL;
rank = ompi_comm_rank ( c->hier_comm );
printf("%d: Dump of hier-struct for comm %s cid %u\n",
rank, c->hier_comm->c_name, c->hier_comm->c_contextid);
printf("%d: No of llead communicators: %d No of lleaders: %d\n",
rank, opal_pointer_array_get_size ( &(c->hier_llead)),
c->hier_num_lleaders );
for ( i=0; i < opal_pointer_array_get_size(&(c->hier_llead)); i++ ) {
current = (mca_coll_hierarch_llead_t*)opal_pointer_array_get_item (&(c->hier_llead), i);
if ( current == NULL ) {
continue;
}
printf("%d: my_leader %d am_leader %d\n", rank,
current->my_lleader, current->am_lleader );
for (j=0; j<c->hier_num_lleaders; j++ ) {
printf("%d: lleader[%d] = %d\n", rank, j, current->lleaders[j]);
}
}
return;
}
int mca_coll_hierarch_ft_event(int state) {
if(OPAL_CRS_CHECKPOINT == state) {
;
}
else if(OPAL_CRS_CONTINUE == state) {
;
}
else if(OPAL_CRS_RESTART == state) {
;
}
else if(OPAL_CRS_TERM == state ) {
;
}
else {
;
}
return OMPI_SUCCESS;
}