ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
ed78553512
openmpi/ompi/mca/bcol/ptpcoll/bcol_ptpcoll_reduce.c

406 строки
16 KiB
C
Исходник Обычный вид История

coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2009-2013 Oak Ridge National Laboratory. All rights reserved.
* Copyright (c) 2009-2012 Mellanox Technologies. All rights reserved.
* Copyright (c) 2013 Los Alamos National Security, LLC. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "ompi/include/ompi/constants.h"
#include "ompi/mca/bcol/bcol.h"
#include "bcol_ptpcoll_reduce.h"
#include "bcol_ptpcoll_utils.h"
static int bcol_ptpcoll_reduce_narray_progress(bcol_function_args_t *input_args,
Fixing broken dependency between ML/BCOLS This is hot-fix patch for the issue reported by Ralph. In future we plan to restructure ml data structure layout. Tested by Nathan. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30619. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-07 23:15:45 +04:00
struct mca_bcol_base_function_t *const_args);
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
static int bcol_ptpcoll_reduce_narray(bcol_function_args_t *input_args,
Fixing broken dependency between ML/BCOLS This is hot-fix patch for the issue reported by Ralph. In future we plan to restructure ml data structure layout. Tested by Nathan. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30619. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-07 23:15:45 +04:00
struct mca_bcol_base_function_t *const_args);
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
#define NARRAY_RECV_NB(narray_node, process_shift, group_size, \
recv_buffer, pack_len, tag, comm, recv_requests, \
num_pending_recvs) \
do { \
int n, rc = OMPI_SUCCESS; \
int dst; \
int comm_dst; \
int offset = 0 ; \
\
/* Recieve data from all relevant childrens */ \
for (n = 0; n < narray_node->n_children; n++) { \
\
dst = narray_node->children_ranks[n] + process_shift; \
if (dst >= group_size) { \
dst -= group_size; \
} \
comm_dst = group_list[dst]; \
\
/* Non blocking send .... */ \
PTPCOLL_VERBOSE(1 , ("Reduce, Irecv data to %d[%d], count %d, tag %d, addr %p", \
dst, comm_dst, pack_len, tag, \
data_buffer)); \
rc = MCA_PML_CALL(irecv((void *)((unsigned char*)recv_buffer + offset), pack_len, MPI_BYTE, \
comm_dst, tag, comm, \
&(recv_requests[*num_pending_recvs]))); \
if( OMPI_SUCCESS != rc ) { \
PTPCOLL_VERBOSE(10, ("Failed to start non-blocking receive")); \
return OMPI_ERROR; \
} \
++(*num_pending_recvs); \
offset += pack_len; \
} \
} while(0)
static inline int narray_reduce(void *data_buffer, void *recv_buffer,
int nrecvs, int count,
struct ompi_datatype_t *dtype, struct ompi_op_t *op,
int *reduction_status) {
int pack_len = count * dtype->super.size;
int i = 0;
void *source_buffer = NULL, *result_buffer = NULL;
source_buffer = data_buffer;
result_buffer = recv_buffer;
for (i = 0; i < nrecvs; i++) {
ompi_op_reduce(op, (void*)((unsigned char*) source_buffer) ,
(void*)((unsigned char*) result_buffer),
count,dtype);
source_buffer = (void *)((unsigned char*)recv_buffer
+ (i+1) * pack_len);
}
*reduction_status = 1;
return OMPI_SUCCESS;
}
static int bcol_ptpcoll_reduce_narray_progress(bcol_function_args_t *input_args,
Fixing broken dependency between ML/BCOLS This is hot-fix patch for the issue reported by Ralph. In future we plan to restructure ml data structure layout. Tested by Nathan. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30619. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-07 23:15:45 +04:00
struct mca_bcol_base_function_t *const_args)
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
{
mca_bcol_ptpcoll_module_t *ptpcoll_module = (mca_bcol_ptpcoll_module_t *)const_args->bcol_module;
int tag = -1;
int rc;
int group_size = ptpcoll_module->group_size;
int *group_list = ptpcoll_module->super.sbgp_partner_module->group_list;
uint32_t buffer_index = input_args->buffer_index;
struct ompi_op_t *op = input_args->op;
ompi_communicator_t* comm = ptpcoll_module->super.sbgp_partner_module->group_comm;
ompi_request_t **send_request =
&ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].requests[0];
ompi_request_t **recv_requests =
&ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].requests[1];
void *data_buffer = NULL;
void *src_buffer = (void *) (
(unsigned char *)input_args->sbuf +
(size_t)input_args->sbuf_offset);
void *recv_buffer = (void *) (
(unsigned char *)input_args->rbuf +
(size_t)input_args->rbuf_offset);
int count = input_args->count;
struct ompi_datatype_t *dtype = input_args->dtype;
int pack_len = input_args->count * input_args->dtype->super.size;
int *active_requests =
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].active_requests);
int matched = false;
int my_group_index = ptpcoll_module->super.sbgp_partner_module->my_index;
int relative_group_index = 0;
netpatterns_tree_node_t *narray_node = NULL;
bool not_sent = false;
int parent_rank = -1, comm_parent_rank = -1;
int group_root_index = input_args->root;
if (!ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].reduce_init_called) {
bcol_ptpcoll_reduce_narray(input_args, const_args);
}
/*
* By default the src buffer is the data buffer,
* only after reduction, the recv buffer becomes the
* data buffer
*/
data_buffer = src_buffer;
relative_group_index = my_group_index - group_root_index;
if (relative_group_index < 0) {
relative_group_index +=group_size;
}
/* keep tag within the limit support by the pml */
tag = (PTPCOLL_TAG_OFFSET + input_args->sequence_num * PTPCOLL_TAG_FACTOR) & (ptpcoll_module->tag_mask);
/* mark this as a collective tag, to avoid conflict with user-level tags */
tag = -tag;
narray_node = &ptpcoll_module->narray_node[relative_group_index];
PTPCOLL_VERBOSE(3, ("reduce, Narray tree Progress"));
PTPCOLL_VERBOSE(8, ("bcol_ptpcoll_reduce_narray, buffer index: %d "
"tag: %d "
"tag_mask: %d "
"sn: %d "
"root: %d [%d]"
"buff: %p ",
buffer_index, tag,
ptpcoll_module->tag_mask, input_args->sequence_num,
input_args->root_flag, input_args->root_route->rank,
data_buffer));
/*
Check if the data was received
*/
if (0 != *active_requests) {
matched = mca_bcol_ptpcoll_test_all_for_match
(active_requests, recv_requests, &rc);
if (OMPI_SUCCESS != rc) {
return OMPI_ERROR;
}
/* All data was received, then do a reduction*/
if(matched) {
narray_reduce(data_buffer, recv_buffer, narray_node->n_children, count, dtype, op,
&ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].reduction_status);
/*
* The reduction result is in the recv buffer, so it is the new data
* buffer
*/
data_buffer = recv_buffer;
/* If not reduced, means also, you might not posted a send */
not_sent = true;
} else {
PTPCOLL_VERBOSE(10, ("reduce root is started"));
return BCOL_FN_STARTED;
}
}
/* I'm root, I'm done */
if (input_args->root_flag) {
return BCOL_FN_COMPLETE;
}
PTPCOLL_VERBOSE(1,("Testing Sending Match"));
/* If send was not posted */
/* Manju: Leaf node should never post in the progress logic */
if (not_sent) {
parent_rank =
ptpcoll_module->narray_node[relative_group_index].parent_rank +
group_root_index;
if (parent_rank >= group_size) {
parent_rank -= group_size;
}
comm_parent_rank = group_list[parent_rank];
PTPCOLL_VERBOSE(1,("Sending data to %d ",comm_parent_rank));
rc = MCA_PML_CALL(isend(data_buffer, pack_len, MPI_BYTE,
comm_parent_rank,
tag, MCA_PML_BASE_SEND_STANDARD, comm, send_request));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to send data"));
return OMPI_ERROR;
}
}
if (0 == mca_bcol_ptpcoll_test_for_match(send_request, &rc)) {
PTPCOLL_VERBOSE(10, ("Test was not matched - %d", rc));
/* Data has not been sent. Return that the collective has been stated
* because we MUST call test on this request once it is finished to
* ensure that it is properly freed. */
return (OMPI_SUCCESS != rc) ? rc : BCOL_FN_STARTED;
}
return BCOL_FN_COMPLETE;
}
static int bcol_ptpcoll_reduce_narray(bcol_function_args_t *input_args,
Fixing broken dependency between ML/BCOLS This is hot-fix patch for the issue reported by Ralph. In future we plan to restructure ml data structure layout. Tested by Nathan. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30619. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-07 23:15:45 +04:00
struct mca_bcol_base_function_t *const_args)
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
{
mca_bcol_ptpcoll_module_t *ptpcoll_module = (mca_bcol_ptpcoll_module_t *)const_args->bcol_module;
int tag;
int rc;
int group_size = ptpcoll_module->group_size;
int *group_list = ptpcoll_module->super.sbgp_partner_module->group_list;
uint32_t buffer_index = input_args->buffer_index;
struct ompi_op_t *op = input_args->op;
ompi_communicator_t* comm = ptpcoll_module->super.sbgp_partner_module->group_comm;
ompi_request_t **recv_requests =
&ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].requests[1];
ompi_request_t **send_request =
&ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].requests[0];
void *data_buffer = NULL;
void *src_buffer = (void *) (
(unsigned char *)input_args->sbuf +
(size_t)input_args->sbuf_offset);
void *recv_buffer = (void *) (
(unsigned char *)input_args->rbuf +
(size_t)input_args->rbuf_offset);
int count = input_args->count;
struct ompi_datatype_t *dtype = input_args->dtype;
int pack_len = input_args->count * input_args->dtype->super.size;
int *active_requests =
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].active_requests);
int matched = true;
int my_group_index = ptpcoll_module->super.sbgp_partner_module->my_index;
int group_root_index = -1;
int relative_group_index = 0;
netpatterns_tree_node_t *narray_node = NULL;
int parent_rank = -1, comm_parent_rank = -1;
/* This is first function call that should be called, not progress.
* The fragmentation code does this, so switch from progress to here.
* The flag indicates whether, we have entered this code *
*/
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].reduce_init_called = true;
PTPCOLL_VERBOSE(1, ("Reduce, Narray tree"));
/* reset active request counter */
(*active_requests) = 0;
/* keep tag within the limit support by the pml */
tag = (PTPCOLL_TAG_OFFSET + input_args->sequence_num * PTPCOLL_TAG_FACTOR) & (ptpcoll_module->tag_mask);
/* mark this as a collective tag, to avoid conflict with user-level flags */
tag = -tag;
PTPCOLL_VERBOSE(1, ("bcol_ptpcoll_reduce_narray, buffer index: %d "
"tag: %d "
"tag_mask: %d "
"sn: %d "
"root: %d "
"buff: %p ",
buffer_index, tag,
ptpcoll_module->tag_mask, input_args->sequence_num,
input_args->root_flag,
src_buffer));
/* Compute Root Index Shift */
group_root_index = input_args->root;
relative_group_index = my_group_index - group_root_index;
if (relative_group_index < 0) {
relative_group_index += group_size;
}
narray_node = &ptpcoll_module->narray_node[relative_group_index];
if (0 == narray_node->n_children) {
PTPCOLL_VERBOSE(10, ("I'm leaf of the data"));
/*
* I'm root of the operation
* send data to N childrens
*/
data_buffer = src_buffer;
goto NARRAY_SEND_DATA;
}
/* Not leaf, either an internal node or root */
NARRAY_RECV_NB(narray_node, group_root_index, group_size,
recv_buffer, pack_len, tag, comm, recv_requests,
active_requests);
/* We have not done reduction, yet */
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].reduction_status = 0;
/* We can not block. So run couple of test for data arrival */
matched = mca_bcol_ptpcoll_test_all_for_match
(active_requests, recv_requests, &rc);
/* Check if received the data */
if(matched) {
narray_reduce(src_buffer, recv_buffer, narray_node->n_children,
count, dtype, op, &ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].reduction_status);
PTPCOLL_VERBOSE(1, ("Reduce, received data from all childrend "));
data_buffer = recv_buffer;
} else {
PTPCOLL_VERBOSE(1, ("reduce root is started"));
return BCOL_FN_STARTED;
}
/* I'm root, I'm done */
if (input_args->root_flag) {
return BCOL_FN_COMPLETE;
}
NARRAY_SEND_DATA:
/*
* Send the data (reduce in case of internal nodes, or just data in
* case of leaf nodes) to the parent
*/
narray_node = &ptpcoll_module->narray_node[relative_group_index];
parent_rank =
ptpcoll_module->narray_node[relative_group_index].parent_rank +
group_root_index;
if (parent_rank >= group_size) {
parent_rank -= group_size;
}
comm_parent_rank = group_list[parent_rank];
PTPCOLL_VERBOSE(1,("Sending data to %d ",comm_parent_rank));
rc = MCA_PML_CALL(isend(data_buffer, pack_len, MPI_BYTE,
comm_parent_rank,
tag, MCA_PML_BASE_SEND_STANDARD, comm, send_request));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to send data"));
return OMPI_ERROR;
}
/* We can not block. So run couple of test for data arrival */
if (0 == mca_bcol_ptpcoll_test_for_match(send_request, &rc)) {
PTPCOLL_VERBOSE(10, ("Test was not matched - %d", rc));
/* No data was received, return no match error */
return (OMPI_SUCCESS != rc) ? rc : BCOL_FN_STARTED;
}
return BCOL_FN_COMPLETE;
}
int bcol_ptpcoll_reduce_init(mca_bcol_base_module_t *super)
{
mca_bcol_base_coll_fn_comm_attributes_t comm_attribs;
mca_bcol_base_coll_fn_invoke_attributes_t inv_attribs;
PTPCOLL_VERBOSE(1,("Initialization Reduce - Narray"));
comm_attribs.bcoll_type = BCOL_REDUCE;
comm_attribs.comm_size_min = 0;
comm_attribs.comm_size_max = 1024 * 1024;
comm_attribs.waiting_semantics = NON_BLOCKING;
inv_attribs.bcol_msg_min = 0;
inv_attribs.bcol_msg_max = 20000; /* range 1 */
inv_attribs.datatype_bitmap = 0xffffffff;
inv_attribs.op_types_bitmap = 0xffffffff;
comm_attribs.data_src = DATA_SRC_KNOWN;
mca_bcol_base_set_attributes(super, &comm_attribs, &inv_attribs,
bcol_ptpcoll_reduce_narray,
bcol_ptpcoll_reduce_narray_progress);
comm_attribs.data_src = DATA_SRC_KNOWN;
return OMPI_SUCCESS;
}
Копировать постоянную ссылку