ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
b68d66bb9b
openmpi/ompi/mca/bcol/ptpcoll/bcol_ptpcoll_allreduce.c

1030 строки
41 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_allreduce.h"
/*
* Recursive K-ing allreduce
*/
static inline int bcol_ptpcoll_allreduce_narray_schedule_extra_node_exchange (mca_bcol_ptpcoll_module_t *ptpcoll_module, netpatterns_k_exchange_node_t *k_node,
void *data_buffer, size_t data_size, ompi_request_t **requests, int *active_requests,
int tag)
{
ompi_communicator_t *comm = ptpcoll_module->super.sbgp_partner_module->group_comm;
int peer_comm_rank, k, offset, rc;
if (EXCHANGE_NODE == k_node->node_type) {
/* the send data resides in the first part of the buffer */
for (k = 0, offset = data_size ; k < k_node->n_extra_sources ; ++k, offset += data_size) {
peer_comm_rank = ptpcoll_module->super.sbgp_partner_module->group_list[k_node->rank_extra_sources_array[k]];
PTPCOLL_VERBOSE(10, ("Recv data from %d, addr %p len %d tag %d",
peer_comm_rank, data_buffer, data_size, tag));
rc = MCA_PML_CALL(irecv((void *)((unsigned char *)data_buffer + offset),
data_size, MPI_BYTE, peer_comm_rank, tag, comm,
&requests[*active_requests]));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to receive data"));
return OMPI_ERROR;
}
++(*active_requests);
}
} else {
peer_comm_rank = ptpcoll_module->super.sbgp_partner_module->group_list[k_node->rank_extra_sources_array[0]];
PTPCOLL_VERBOSE(10, ("Send data to %d, addr %p len %d tag %d",
peer_comm_rank, data_buffer, data_size, tag));
rc = MCA_PML_CALL(isend(data_buffer, data_size, MPI_BYTE, peer_comm_rank,
tag, MCA_PML_BASE_SEND_STANDARD, comm,
&(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to send data"));
return OMPI_ERROR;
}
++(*active_requests);
}
return OMPI_SUCCESS;
}
static inline void bcol_ptpcoll_allreduce_narray_reduce (void *data_buffer, struct ompi_datatype_t *data_type, int count, struct ompi_op_t *op, int sources)
{
size_t data_size = mca_bcol_base_get_buff_length(data_type, count);
for (int k = 0, offset = data_size ; k < sources ; ++k, offset += data_size) {
ompi_op_reduce(op, (char *) data_buffer + offset, data_buffer, count, data_type);
}
}
static int bcol_ptpcoll_allreduce_narraying_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;
void *data_buffer = (void *) ( (unsigned char *) input_args->sbuf +
(size_t) input_args->sbuf_offset);
struct ompi_datatype_t *data_type = input_args->dtype;
uint32_t buffer_index = input_args->buffer_index;
struct ompi_op_t *op = input_args->op;
int count = input_args->count;
int *active_requests =
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].active_requests);
int tag = ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].tag - 1;
int *group_list = ptpcoll_module->super.sbgp_partner_module->group_list;
int k, rc, peer, group_peer;
int offset = 0;
ompi_communicator_t *comm = ptpcoll_module->super.sbgp_partner_module->group_comm;
ompi_request_t **requests =
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].requests;
netpatterns_k_exchange_node_t *k_node = &ptpcoll_module->knomial_exchange_tree;
int k_radix = k_node->tree_order;
size_t data_size = mca_bcol_base_get_buff_length(data_type, count);
int *iteration =
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].iteration);
/* if we are just staring the collective and there are extra sources then schedule the
* extra node exchange. otherwise check if the exchange is complete. */
if (-1 == *iteration) {
if (0 < k_node->n_extra_sources) {
if (!(*active_requests)) {
rc = bcol_ptpcoll_allreduce_narray_schedule_extra_node_exchange (ptpcoll_module, k_node, data_buffer, data_size,
requests, active_requests, tag);
if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
return rc;
}
}
/* check for extra node exchange completion */
if (!mca_bcol_ptpcoll_test_all_for_match (active_requests, requests, &rc)) {
return (OMPI_SUCCESS == rc) ? BCOL_FN_STARTED : rc;
}
if (EXCHANGE_NODE == k_node->node_type) {
bcol_ptpcoll_allreduce_narray_reduce (data_buffer, data_type, count, op, k_node->n_extra_sources);
}
}
/* start recursive k-ing */
*iteration = 0;
}
if (*iteration < k_node->n_exchanges) {
if (*active_requests) {
if (!mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc)) {
return (OMPI_SUCCESS == rc) ? BCOL_FN_STARTED : rc;
}
++(*iteration);
bcol_ptpcoll_allreduce_narray_reduce (data_buffer, data_type, count, op, k_radix - 1);
}
}
for ( ; *iteration < k_node->n_exchanges ; ++(*iteration)) {
for (k = 0; k < k_radix - 1; k++) {
group_peer = k_node->rank_exchanges[*iteration][k];
peer = group_list[group_peer];
PTPCOLL_VERBOSE(10, ("Send data to %d, addr %p len %d tag %d",
peer, data_buffer, data_size, tag));
rc = MCA_PML_CALL(isend(data_buffer, data_size, MPI_BYTE, peer, tag,
MCA_PML_BASE_SEND_STANDARD, comm,
&(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to send data"));
return OMPI_ERROR;
}
++(*active_requests);
}
for (k = 0, offset = data_size ; k < k_radix - 1 ; ++k, offset += data_size) {
group_peer = k_node->rank_exchanges[*iteration][k];
peer = group_list[group_peer];
PTPCOLL_VERBOSE(10, ("Recv data from %d, addr %p len %d tag %d",
peer, data_buffer, data_size, tag));
rc = MCA_PML_CALL(irecv((void *)((unsigned char *)data_buffer + offset ),
data_size, MPI_BYTE, peer, tag, comm,
&requests[*active_requests]));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to receive data"));
return OMPI_ERROR;
}
++(*active_requests);
}
if (!mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc)) {
return (OMPI_SUCCESS == rc) ? BCOL_FN_STARTED : rc;
}
bcol_ptpcoll_allreduce_narray_reduce (data_buffer, data_type, count, op, k_radix - 1);
}
/* ensure extra nodes get the result */
if (0 < k_node->n_extra_sources) {
if (!(*active_requests)) {
int peer_comm_rank;
if (EXTRA_NODE == k_node->node_type) {
peer_comm_rank = ptpcoll_module->super.sbgp_partner_module->group_list[k_node->rank_extra_sources_array[0]];
PTPCOLL_VERBOSE(10, ("EXTRA_NODE: Recv data from %d, addr %p len %d tag %d",
peer_comm_rank, data_buffer, data_size, tag));
rc = MCA_PML_CALL(irecv(data_buffer, data_size, MPI_BYTE, peer_comm_rank,
tag, comm, &requests[*active_requests]));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to receive data"));
return OMPI_ERROR;
}
++(*active_requests);
} else {
for (k = 0; k < k_node->n_extra_sources; k++) {
peer_comm_rank = ptpcoll_module->super.sbgp_partner_module->group_list[k_node->rank_extra_sources_array[k]];
PTPCOLL_VERBOSE(10, ("EXCHANGE_NODE: Send data to %d, addr %p len %d tag %d",
peer_comm_rank, data_buffer, data_size, tag));
rc = MCA_PML_CALL(isend(data_buffer, data_size, MPI_BYTE, peer_comm_rank,
tag, MCA_PML_BASE_SEND_STANDARD, comm,
&(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to send data"));
return OMPI_ERROR;
}
++(*active_requests);
}
}
}
if (!mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc)) {
return (OMPI_SUCCESS == rc) ? BCOL_FN_STARTED : rc;
}
}
return BCOL_FN_COMPLETE;
}
int bcol_ptpcoll_allreduce_narraying_init(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;
uint64_t sequence_number = input_args->sequence_num;
uint32_t buffer_index = input_args->buffer_index;
int count = input_args->count;
struct ompi_datatype_t *dtype = input_args->dtype;
size_t buffer_size;
int tag;
tag = (PTPCOLL_TAG_OFFSET + sequence_number * PTPCOLL_TAG_FACTOR) & (ptpcoll_module->tag_mask);
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].tag = tag = -tag;
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].radix_mask = 1;
/* start with extra node exchange if needed */
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].iteration = -1;
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].active_requests = 0;
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].status = PTPCOLL_NOT_STARTED;
/*
* ML bufer is segmented into k segments and each of the k segment is used
* for reductions
*/
/* This has to be based on ml buffer size. Need to take into account the space used
* by the headers of other bcol modules. */
buffer_size = ptpcoll_module->ml_mem.size_buffer - BCOL_HEADER_MAX;
assert(buffer_size >= count * dtype->super.size *
ptpcoll_module->k_nomial_radix);
return bcol_ptpcoll_allreduce_narraying_progress (input_args, const_args);
}
static inline int compute_seg_index(int peer, int kpow_num, int tree_order) {
int peer_base, peer_position, peer_base_rank, peer_index;
peer_base = peer / (kpow_num * tree_order);
peer_base_rank = peer_base * kpow_num * tree_order ;
peer_position = peer_base_rank == 0 ? peer : peer % (peer_base_rank);
peer_index = peer_position / kpow_num ;
return peer_index;
}
int compute_knomial_allgather_offsets(int group_index, int count, struct
ompi_datatype_t *dtype,int k_radix,int n_exchanges,
int **offsets){
int modulo_group_size;
size_t seg_count, seg_size, seg_index, seg_offset;
size_t block_offset, block_count;
int exchange_step;
ptrdiff_t lb, extent;
if (0 >= n_exchanges) {
PTPCOLL_VERBOSE(10,("Nothing to initialize "));
return 0;
}
modulo_group_size = 1;
seg_count = count / k_radix;
ompi_datatype_get_extent(dtype, &lb, &extent);
seg_size = seg_count * extent;
seg_index = group_index % k_radix;
seg_offset = seg_index * seg_size;
offsets[0][BLOCK_OFFSET] = block_offset = 0;
offsets[0][BLOCK_COUNT] = block_count = count;
offsets[0][LOCAL_REDUCE_SEG_OFFSET] = seg_offset;
offsets[0][SEG_SIZE] = seg_size;
for(exchange_step = 1; exchange_step < n_exchanges; exchange_step++) {
/* Previous step's segment is this exchange step's block */
block_count = seg_count;
block_offset = seg_offset;
/* Divide the segment into k parts */
seg_count = seg_count / k_radix;
seg_size = seg_count * extent;
/* Among different segments in block, which segment should I reduce ? */
/* For allgather phase, I will not send out this segment to peers */
modulo_group_size *= k_radix;
seg_index = compute_seg_index(group_index, modulo_group_size, k_radix);
seg_offset = seg_index * seg_size;
offsets[exchange_step][BLOCK_OFFSET] = block_offset;
offsets[exchange_step][LOCAL_REDUCE_SEG_OFFSET] = seg_offset;
offsets[exchange_step][BLOCK_COUNT] = block_count;
offsets[exchange_step][SEG_SIZE] = seg_size;
/* Change to absolute offset */
seg_offset = block_offset + seg_offset;
}
return 0;
}
static inline int compute_send_segment_size(int block_offset,
int send_offset,
int segment_size,
int padded_offset) {
int send_size = -1;
/* segment to be sent starts here */
int segment_offset = block_offset + send_offset ;
send_size = (segment_offset + segment_size) >= padded_offset ?
segment_size - (segment_offset + segment_size - padded_offset) : segment_size;
return send_size;
}
static inline int compute_recv_segment_size(int block_offset,
int recv_offset,
int segment_size,
int padded_offset) {
int recv_size = -1;
/* segment to be sent starts here */
int segment_offset = block_offset + recv_offset ;
recv_size = (segment_offset + segment_size) >= padded_offset ?
segment_size - (segment_offset + segment_size - padded_offset) : segment_size;
return recv_size;
}
/*
*
* K-nomial Reduce Scatter
* Example k=3 n=9
*
* | ABCDEFGH |0|
*
* Number of Exchange steps = log (basek) n
* Number of steps in exchange step = k (radix)
*
* block_size = Size of data that is reduce in exchange step
* segment_size = Size of data that is send or received by rank in radix step
*
* block_size = segment_size * k
*
* my_block_start_addr = Address of the segment in the block where I reference my
* offsets
*
* This is version 1 : Experimenting with decoupling offset calcuations
*/
int bcol_ptpcoll_allreduce_recursivek_scatter_reduce(mca_bcol_ptpcoll_module_t *ptpcoll_module,
const int buffer_index, void *sbuf,
void *rbuf,
struct ompi_op_t *op,
const int count, struct ompi_datatype_t *dtype,
const int relative_group_index,
const int padded_start_byte){
int blocks_in_step =
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].radix_mask;
int tag = ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].tag - 1;
int *group_list = ptpcoll_module->super.sbgp_partner_module->group_list;
netpatterns_k_exchange_node_t *k_node = &ptpcoll_module->knomial_exchange_tree;
mca_bcol_ptpcoll_component_t *cm =
&mca_bcol_ptpcoll_component;
void *my_block_start_addr = NULL, *my_block_addr = NULL;
int i, k, group_peer, peer ;
int k_radix = k_node->tree_order;
int rc = OMPI_SUCCESS;
ompi_communicator_t* comm = ptpcoll_module->super.sbgp_partner_module->group_comm;
ompi_request_t **requests =
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].requests;
int *active_requests =
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].active_requests);
int completed;
void *my_recv_start_addr, *my_recv_addr;
size_t block_offset, reduce_seg_offset, send_offset, recv_offset;
int seg_size, block_size;
int block_count, seg_count;
ptrdiff_t lb, extent;
ompi_datatype_get_extent(dtype, &lb, &extent);
my_recv_start_addr = rbuf;
my_block_start_addr = sbuf;
block_count = count;
block_size = count * extent;
for (i = ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].iteration;
i < k_node->n_exchanges; i++, blocks_in_step *= cm->narray_knomial_radix) {
block_offset = ptpcoll_module->allgather_offsets[i][BLOCK_OFFSET];
reduce_seg_offset = ptpcoll_module->allgather_offsets[i][LOCAL_REDUCE_SEG_OFFSET];
block_count = ptpcoll_module->allgather_offsets[i][BLOCK_COUNT];
seg_size = ptpcoll_module->allgather_offsets[i][SEG_SIZE];
block_size = block_count * extent;
PTPCOLL_VERBOSE(10,("Block offset %d, reduce_seg_offset %d, block_count %d seg_size %d",
block_offset, reduce_seg_offset, block_count, seg_size));
seg_count = block_count / k_radix;
my_block_addr = (void*)((char*)my_block_start_addr + block_offset);
my_recv_addr = (void*)((char*)my_recv_start_addr + block_offset);
for (k = 0; k < k_radix - 1; k++) {
size_t soffset;
int snd_size = 0;
group_peer = k_node->rank_exchanges[i][k];
peer = group_list[group_peer];
send_offset = reduce_seg_offset + (seg_size * (k + 1));
if ((int)send_offset + seg_size > block_size) {
send_offset = send_offset % block_size;
}
PTPCOLL_VERBOSE(10, ("Send data to %d,send offset %d len %d",
peer, send_offset, seg_size));
soffset = send_offset;
snd_size =
compute_send_segment_size((int)block_offset,(int)soffset,(int)seg_size,padded_start_byte);
if (snd_size > 0) {
rc = MCA_PML_CALL(isend((void *)((unsigned char *)my_block_addr
+ soffset),
snd_size, MPI_BYTE,
peer, tag, MCA_PML_BASE_SEND_STANDARD, comm,
&(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to send the segment to %d", peer));
return OMPI_ERROR;
}
++(*active_requests);
}
}
/*
* Receive the segments to tmp addr and then do a reduction
*/
for (k = 0; k < k_radix - 1; k++) {
int recv_size=0;
group_peer = k_node->rank_exchanges[i][k];
peer = group_list[group_peer];
recv_offset = reduce_seg_offset + (seg_size * (k+1));
if ((int)recv_offset + seg_size > block_size) {
recv_offset = recv_offset % block_size;
}
PTPCOLL_VERBOSE(10, ("Receive data to receive buffer at offset %d\n",
recv_offset));
recv_size = compute_recv_segment_size((int)block_offset,
(int)reduce_seg_offset, (int)seg_size,
padded_start_byte);
if (recv_size > 0 ) {
rc = MCA_PML_CALL(irecv((void *)((unsigned char *)
my_recv_addr + recv_offset),
recv_size, MPI_BYTE,
peer, tag, comm, &requests[*active_requests]));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to receive the segment from %d", peer));
return OMPI_ERROR;
}
++(*active_requests);
}
}
completed = 0;
while(!completed){
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
}
/* Do a reduction on received buffers */
{
void *src_data_buffer = NULL, *dst_data_buffer = NULL;
int reduce_data_count = 0;
src_data_buffer = my_block_addr;
dst_data_buffer = my_recv_addr;
for (k = 0; k < k_radix - 1; k++) {
recv_offset = reduce_seg_offset + (seg_size * (k+1));
if ((int)recv_offset + seg_size > block_size) {
recv_offset = recv_offset % block_size;
}
reduce_data_count = (int)(block_offset + reduce_seg_offset) + seg_size >= padded_start_byte ?
(seg_size - (((int)(block_offset + reduce_seg_offset) + seg_size) - padded_start_byte))/(int)dtype->super.size
: (int)seg_count;
if (reduce_data_count > 0) {
ompi_3buff_op_reduce(op,
(void*)((unsigned char*)my_recv_addr + recv_offset),
(void*)((unsigned char*)src_data_buffer +
reduce_seg_offset),
(void*)((unsigned char*)dst_data_buffer +
reduce_seg_offset),
reduce_data_count,dtype);
}
src_data_buffer = dst_data_buffer;
}
}
/* After first iteration we have data (to work with) in recv buffer */
my_block_start_addr = rbuf;
}
return rc;
}
int bcol_ptpcoll_allreduce_knomial_allgather(mca_bcol_ptpcoll_module_t *ptpcoll_module,
const int buffer_index,
void *sbuf,void *rbuf, int count, struct
ompi_datatype_t *dtype,
const int relative_group_index,
const int padded_start_byte){
size_t block_offset = 0, send_offset = 0, recv_offset = 0;
int seg_size=0, block_size=0;
int i,k,completed;
void *my_block_start_addr = rbuf, *my_block_addr;
size_t block_count = count;
netpatterns_k_exchange_node_t *k_node = &ptpcoll_module->knomial_exchange_tree;
int k_radix = k_node->tree_order;
int peer, group_peer;
int rc = OMPI_SUCCESS;
int tag = ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].tag - 1;
int *active_requests =
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].active_requests);
ompi_request_t **requests =
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].requests;
ompi_communicator_t* comm = ptpcoll_module->super.sbgp_partner_module->group_comm;
int exchange_step;
int *group_list = ptpcoll_module->super.sbgp_partner_module->group_list;
ptrdiff_t lb, extent;
ompi_datatype_get_extent(dtype, &lb, &extent);
for (i = ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].iteration;
i < k_node->n_exchanges; i++) {
exchange_step = k_node->n_exchanges - 1 - i;
block_offset = ptpcoll_module->allgather_offsets[exchange_step][BLOCK_OFFSET];
send_offset = ptpcoll_module->allgather_offsets[exchange_step][LOCAL_REDUCE_SEG_OFFSET];
block_count = ptpcoll_module->allgather_offsets[exchange_step][BLOCK_COUNT];
seg_size = ptpcoll_module->allgather_offsets[exchange_step][SEG_SIZE];
block_size = block_count * extent;
PTPCOLL_VERBOSE(10, ("Send offset %d block_offset %d seg_size %\n",
send_offset, block_offset, seg_size));
my_block_addr = (void*)((unsigned char*)my_block_start_addr + block_offset);
for (k = 0; k < k_radix - 1; k++) {
size_t soffset=0; int snd_size = 0;
group_peer = k_node->rank_exchanges[exchange_step][k];
peer = group_list[group_peer];
soffset = send_offset;
snd_size = compute_send_segment_size((int)block_offset,
(int)soffset,
(int)seg_size,
padded_start_byte);
if (snd_size > 0) {
rc = MCA_PML_CALL(isend((void *)((unsigned char *)my_block_addr
+ soffset),
snd_size, MPI_BYTE,
peer, tag, MCA_PML_BASE_SEND_STANDARD, comm,
&(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to send the segment to %d", peer));
return OMPI_ERROR;
}
++(*active_requests);
}
PTPCOLL_VERBOSE(10, ("Send data to receive buffer at offset %d to %d\n",
send_offset, peer));
}
for (k = 0; k < k_radix - 1; k++) {
int recv_size=0;
group_peer = k_node->rank_exchanges[exchange_step][k];
peer = group_list[group_peer];
recv_offset = send_offset + (k + 1) * seg_size;
if ((int)recv_offset + seg_size > block_size){
recv_offset = recv_offset % block_size;
}
PTPCOLL_VERBOSE(10, ("Receive data to receive buffer at offset %d from %d\n",
recv_offset, peer));
recv_size = compute_recv_segment_size((int)block_offset,
(int)recv_offset,
(int)seg_size,
padded_start_byte);
if (recv_size > 0) {
rc = MCA_PML_CALL(irecv((void *)((unsigned char *)
my_block_addr + recv_offset),
recv_size, MPI_BYTE,
peer, tag, comm, &requests[*active_requests]));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to receive the segment from %d", peer));
return OMPI_ERROR;
}
++(*active_requests);
}
}
completed = 0;
while(!completed){
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
}
block_count = block_count * k_radix;
block_size = block_count * extent;
}
return rc;
}
static inline int compute_padding_count(int count, int k_radix, int n_exchanges){
bool fpadding = false;
size_t dsize;
int i, pad_count=0, kpow;
/* is padding required */
dsize = count;
kpow = 1;
for ( i=0; i < n_exchanges; i++) {
if (dsize % k_radix) {
fpadding = true;
}
dsize /= k_radix;
kpow *= k_radix;
}
if (fpadding) {
pad_count = count % kpow;
pad_count = kpow - pad_count;
}
return pad_count;
}
int bcol_ptpcoll_allreduce_recursivek_scatter_reduce_allgather_init(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;
struct ompi_op_t *op = input_args->op;
int tag;
int my_group_index = ptpcoll_module->super.sbgp_partner_module->my_index;
uint64_t sequence_number = input_args->sequence_num;
uint32_t buffer_index = input_args->buffer_index;
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 *iteration =
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].iteration);
int *active_requests =
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].active_requests);
int *status =
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].status);
ptrdiff_t lb, extent;
/* Get the knomial tree */
netpatterns_k_exchange_node_t *k_node = &ptpcoll_module->knomial_exchange_tree;
int k_radix = k_node->tree_order;
int n_exchanges = k_node->n_exchanges;
int padded_start_byte;
int padding_count = compute_padding_count(count, k_radix, n_exchanges);
ompi_datatype_get_extent(dtype, &lb, &extent);
padded_start_byte = count * extent;
/* Init for making the functions Re-entrant */
tag = (PTPCOLL_TAG_OFFSET + sequence_number * PTPCOLL_TAG_FACTOR) & (ptpcoll_module->tag_mask);
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].tag = tag = -tag;
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].radix_mask = 1;
*active_requests = 0;
*iteration = -1;
*status = PTPCOLL_NOT_STARTED;
*iteration = 0;
compute_knomial_allgather_offsets(my_group_index,count + padding_count, dtype,k_radix,n_exchanges,
ptpcoll_module->allgather_offsets);
/* Perform a recursive k'ing reduce scatter */
bcol_ptpcoll_allreduce_recursivek_scatter_reduce(ptpcoll_module, buffer_index,
src_buffer, recv_buffer, op, count + padding_count, dtype,
my_group_index,padded_start_byte);
/* Perform a recursive k'ing allgather */
bcol_ptpcoll_allreduce_knomial_allgather(ptpcoll_module,
buffer_index,
src_buffer, recv_buffer, count + padding_count, dtype,
my_group_index, padded_start_byte);
return BCOL_FN_COMPLETE;
}
int bcol_ptpcoll_allreduce_recursivek_scatter_reduce_extra(mca_bcol_ptpcoll_module_t *ptpcoll_module,
int buffer_index,
void *sbuf,
void *rbuf,
struct ompi_op_t *op,
const int count, struct ompi_datatype_t *dtype){
int tag = ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].tag - 1;
netpatterns_k_exchange_node_t *k_node = &ptpcoll_module->knomial_exchange_tree;
int k, peer ;
int rc = OMPI_SUCCESS;
ompi_communicator_t* comm = ptpcoll_module->super.sbgp_partner_module->group_comm;
int block_count, block_size;
char *tmprecv_buffer = NULL, *data_src_buffer, *data_dst_buffer;
ptrdiff_t lb, extent;
ompi_datatype_get_extent(dtype, &lb, &extent);
block_count = count;
block_size = count * extent;
if (0 < block_size) {
tmprecv_buffer = (void*)malloc(block_size);
}
data_src_buffer = sbuf;
data_dst_buffer = rbuf;
if (EXCHANGE_NODE == k_node->node_type) {
for (k = 0; k < k_node->n_extra_sources; k++){
peer = ptpcoll_module->super.sbgp_partner_module->group_list[
k_node->rank_extra_sources_array[k]];
rc = MCA_PML_CALL(recv((void *)((unsigned char *)tmprecv_buffer),
block_size, MPI_BYTE,
peer, tag, comm, MPI_STATUS_IGNORE));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to receive the segment from %d", peer));
rc = OMPI_ERROR;
goto clean;
}
ompi_3buff_op_reduce(op, (void*)((unsigned char*)data_src_buffer),
(void*)((unsigned char*)tmprecv_buffer),
(void*)((unsigned char*)data_dst_buffer),
block_count,dtype);
data_src_buffer = data_dst_buffer;
}
} else {
peer = ptpcoll_module->super.sbgp_partner_module->group_list[
k_node->rank_extra_sources_array[0]];
rc = MCA_PML_CALL(send((void *)((unsigned char *)sbuf),
block_size, MPI_BYTE,
peer, tag, MCA_PML_BASE_SEND_STANDARD, comm));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to send data"));
rc = OMPI_ERROR;
goto clean;
}
}
clean:
if (tmprecv_buffer) {
free(tmprecv_buffer);
}
return rc;
}
int bcol_ptpcoll_allreduce_knomial_allgather_extra(mca_bcol_ptpcoll_module_t *ptpcoll_module,
int buffer_index,
void *sbuf,
void *rbuf,
const int count, struct ompi_datatype_t *dtype){
int tag = ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].tag - 1;
netpatterns_k_exchange_node_t *k_node = &ptpcoll_module->knomial_exchange_tree;
int k, peer ;
int rc = OMPI_SUCCESS;
ompi_communicator_t* comm = ptpcoll_module->super.sbgp_partner_module->group_comm;
int block_size, completed;
ompi_request_t **requests =
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].requests;
int *active_requests =
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].active_requests);
ptrdiff_t lb, extent;
ompi_datatype_get_extent(dtype, &lb, &extent);
block_size = count * extent;
if (EXTRA_NODE == k_node->node_type) {
peer = ptpcoll_module->super.sbgp_partner_module->group_list[
k_node->rank_extra_sources_array[0]];
rc = MCA_PML_CALL(irecv((void *)((unsigned char *)rbuf),
block_size, MPI_BYTE,
peer, tag, comm, &requests[*active_requests]));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to receive data"));
return OMPI_ERROR;
}
++(*active_requests);
} else {
for (k = 0; k < k_node->n_extra_sources; k++) {
peer = ptpcoll_module->super.sbgp_partner_module->group_list[
k_node->rank_extra_sources_array[k]];
rc = MCA_PML_CALL(isend((void *)((unsigned char *)rbuf),
block_size, MPI_BYTE,
peer, tag, MCA_PML_BASE_SEND_STANDARD, comm,
&(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to send data"));
return OMPI_ERROR;
}
++(*active_requests);
}
}
completed = 0;
while(!completed){
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
}
return rc;
}
int bcol_ptpcoll_allreduce_recursivek_scatter_reduce_allgather_extra_init(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;
struct ompi_op_t *op = input_args->op;
int tag;
int my_group_index = ptpcoll_module->super.sbgp_partner_module->my_index;
uint64_t sequence_number = input_args->sequence_num;
uint32_t buffer_index = input_args->buffer_index;
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 *iteration =
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].iteration);
int *active_requests =
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].active_requests);
int *status =
&(ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].status);
ptrdiff_t lb, extent;
/* Get the knomial tree */
netpatterns_k_exchange_node_t *k_node = &ptpcoll_module->knomial_exchange_tree;
int k_radix = k_node->tree_order;
int n_exchanges = k_node->n_exchanges;
int padded_start_byte;
int padding_count = compute_padding_count(count, k_radix, n_exchanges);
void *tmpsrc_buffer = NULL;
ompi_datatype_get_extent(dtype, &lb, &extent);
padded_start_byte = count * extent;
/* Init for making the functions Re-entrant */
tag = (PTPCOLL_TAG_OFFSET + sequence_number * PTPCOLL_TAG_FACTOR) & (ptpcoll_module->tag_mask);
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].tag = tag = -tag;
ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].radix_mask = 1;
*active_requests = 0;
*iteration = -1;
*status = PTPCOLL_NOT_STARTED;
*iteration = 0;
compute_knomial_allgather_offsets(my_group_index,count + padding_count, dtype,k_radix,n_exchanges,
ptpcoll_module->allgather_offsets);
if (EXCHANGE_NODE == k_node->node_type) {
bcol_ptpcoll_allreduce_recursivek_scatter_reduce_extra(ptpcoll_module,
buffer_index,
src_buffer, recv_buffer, op, count, dtype);
tmpsrc_buffer = src_buffer;
if ( k_node->n_extra_sources > 0){
tmpsrc_buffer = recv_buffer;
}
bcol_ptpcoll_allreduce_recursivek_scatter_reduce(ptpcoll_module, buffer_index,
tmpsrc_buffer, recv_buffer, op, count + padding_count, dtype,
my_group_index,padded_start_byte);
bcol_ptpcoll_allreduce_knomial_allgather(ptpcoll_module,
buffer_index,
src_buffer, recv_buffer, count + padding_count, dtype,
my_group_index, padded_start_byte);
bcol_ptpcoll_allreduce_knomial_allgather_extra(ptpcoll_module,
buffer_index,
src_buffer, recv_buffer, count, dtype);
}
else if (EXTRA_NODE == k_node->node_type) {
bcol_ptpcoll_allreduce_recursivek_scatter_reduce_extra(ptpcoll_module,
buffer_index,
src_buffer, recv_buffer, op, count, dtype);
bcol_ptpcoll_allreduce_knomial_allgather_extra(ptpcoll_module,
buffer_index,
src_buffer, recv_buffer, count, dtype);
}
return BCOL_FN_COMPLETE;
}
/*
* Register allreduce functions to the BCOL function table,
* so they can be selected
*/
int bcol_ptpcoll_allreduce_init(mca_bcol_base_module_t *super)
{
mca_bcol_ptpcoll_module_t *ptpcoll_module =
(mca_bcol_ptpcoll_module_t *) super;
mca_bcol_base_coll_fn_comm_attributes_t comm_attribs;
mca_bcol_base_coll_fn_invoke_attributes_t inv_attribs;
comm_attribs.bcoll_type = BCOL_ALLREDUCE;
comm_attribs.comm_size_min = 0;
comm_attribs.comm_size_max = 1024 * 1024;
/* not an accurate attribute, none of these algorithms
* are non-blocking
*/
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_allreduce_narraying_init,
bcol_ptpcoll_allreduce_narraying_progress);
inv_attribs.bcol_msg_min = 10000000;
inv_attribs.bcol_msg_max = 10485760; /* range 4 */
if (ptpcoll_module->pow_knum == ptpcoll_module->group_size) {
mca_bcol_base_set_attributes(super, &comm_attribs, &inv_attribs,
bcol_ptpcoll_allreduce_recursivek_scatter_reduce_allgather_init,
NULL);
} else {
mca_bcol_base_set_attributes(super, &comm_attribs, &inv_attribs,
bcol_ptpcoll_allreduce_recursivek_scatter_reduce_allgather_extra_init,
NULL);
}
return OMPI_SUCCESS;
}
Копировать постоянную ссылку