ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
daeb9b6c4f
openmpi/ompi/mca/bcol/ptpcoll/bcol_ptpcoll_allgather.c

606 строки
21 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
/*
* Copyright (c) 2009-2012 Oak Ridge National Laboratory. All rights reserved.
* Copyright (c) 2009-2012 Mellanox Technologies. 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 allgather
*/
/*
*
* Recurssive k-ing algorithm
* Example k=3 n=9
*
*
* Number of Exchange steps = log (basek) n
* Number of steps in exchange step = k (radix)
*
*/
int bcol_ptpcoll_k_nomial_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
{
/* local variables */
mca_bcol_ptpcoll_module_t *ptpcoll_module = (mca_bcol_ptpcoll_module_t *) const_args->bcol_module;
int *group_list = ptpcoll_module->super.sbgp_partner_module->group_list;
netpatterns_k_exchange_node_t *exchange_node = &ptpcoll_module->knomial_allgather_tree;
int my_group_index = ptpcoll_module->super.sbgp_partner_module->my_index;
int group_size = ptpcoll_module->group_size;
int *list_connected = ptpcoll_module->super.list_n_connected; /* critical for hierarchical colls */
int tag;
int i, j;
int knt;
int comm_src, comm_dst, src, dst;
int recv_offset, recv_len;
int send_offset, send_len;
uint32_t buffer_index = input_args->buffer_index;
int pow_k, 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 = 0; /* initialized */
void *data_buffer = (void*)(
(unsigned char *) input_args->sbuf +
(size_t) input_args->sbuf_offset);
int pack_len = input_args->count * input_args->dtype->super.size;
#if 0
fprintf(stderr,"entering p2p allgather pack_len %d. exchange node: %p\n",pack_len, exchange_node);
#endif
/* initialize the iteration counter */
int *iteration = &ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].iteration;
*iteration = 0;
/* reset active request counter */
*active_requests = 0;
/* keep tag within the limit supported 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;
/* k-nomial parameters */
tree_order = exchange_node->tree_order;
pow_k = exchange_node->log_tree_order;
/* let's begin the collective, starting with extra ranks and their
* respective proxies
*/
if( EXTRA_NODE == exchange_node->node_type ) {
/* then I will send to my proxy rank*/
dst = exchange_node->rank_extra_sources_array[0];
/* find rank in the communicator */
comm_dst = group_list[dst];
/* now I need to calculate my own offset */
knt = 0;
for (i = 0 ; i < my_group_index; i++){
knt += list_connected[i];
}
/* send the data to my proxy */
rc = MCA_PML_CALL(isend((void *) ( (unsigned char *) data_buffer +
knt*pack_len),
pack_len * list_connected[my_group_index],
MPI_BYTE,
comm_dst, tag,
MCA_PML_BASE_SEND_STANDARD, comm,
&(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10,("Failed to isend data"));
return OMPI_ERROR;
}
++(*active_requests);
/* now I go ahead and post the receive from my proxy */
comm_src = comm_dst;
knt = 0;
for( i =0; i < group_size; i++){
knt += list_connected[i];
}
rc = MCA_PML_CALL(irecv(data_buffer,
knt * pack_len,
MPI_BYTE,
comm_src,
tag , comm, &(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to post ireceive "));
return OMPI_ERROR;
}
++(*active_requests);
/* poll for completion */
/* this polls internally */
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
if(completed){
/* go to buffer release */
goto FINISHED;
}else{
/* save state and hop out
* nothing to save here
*/
return ((OMPI_SUCCESS != rc) ? OMPI_ERROR : BCOL_FN_STARTED);
}
}else if ( 0 < exchange_node->n_extra_sources ) {
/* I am a proxy for someone */
src = exchange_node->rank_extra_sources_array[0];
/* find the rank in the communicator */
comm_src = group_list[src];
knt = 0;
for(i = 0; i < src; i++){
knt += list_connected[i];
}
/* post the receive */
rc = MCA_PML_CALL(irecv((void *) ( (unsigned char *) data_buffer
+ knt*pack_len),
pack_len * list_connected[src],
MPI_BYTE,
comm_src,
tag , comm, &(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to post ireceive "));
return OMPI_ERROR;
}
++(*active_requests);
/* poll for completion */
/* this routine polls internally */
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
if(!completed){
/* save state and hop out
* We really do need to block here so set
* the iteration to -1 indicating we need to
* finish this part first
*/
*iteration = -1;
return ((OMPI_SUCCESS != rc )? OMPI_ERROR : BCOL_FN_STARTED);
}
}
/* we start the recursive k - ing phase */
/* fprintf(stderr,"tree order %d pow_k %d \n",tree_order,pow_k);*/
for( i = 0; i < pow_k; i++) {
for(j = 0; j < (tree_order - 1); j++) {
/* send phase */
dst = exchange_node->rank_exchanges[i][j];
if( dst < 0 ){
continue;
}
comm_dst = group_list[dst];
send_offset = exchange_node->payload_info[i][j].s_offset * pack_len;
send_len = exchange_node->payload_info[i][j].s_len * pack_len;
/* debug print */
/* fprintf(stderr,"sending %d bytes to rank %d at offset %d\n",send_len, */
/* comm_dst,send_offset); */
rc = MCA_PML_CALL(isend((void*)((unsigned char *) data_buffer +
send_offset),
send_len,
MPI_BYTE,
comm_dst, tag,
MCA_PML_BASE_SEND_STANDARD, comm,
&(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10,("Failed to isend data"));
return OMPI_ERROR;
}
++(*active_requests);
/* sends are posted */
}
/* Now post the recv's */
for( j = 0; j < (tree_order - 1); j++ ) {
/* recv phase */
src = exchange_node->rank_exchanges[i][j];
if( src < 0 ) {
continue;
}
comm_src = group_list[src];
recv_offset = exchange_node->payload_info[i][j].r_offset * pack_len;
recv_len = exchange_node->payload_info[i][j].r_len * pack_len;
/* debug print */
/* fprintf(stderr,"recving %d bytes to rank %d at offset %d\n",recv_len, */
/* comm_src,recv_offset); */
/* post the receive */
rc = MCA_PML_CALL(irecv((void *) ((unsigned char *) data_buffer +
recv_offset),
recv_len,
MPI_BYTE,
comm_src,
tag, comm, &(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to post ireceive "));
return OMPI_ERROR;
}
++(*active_requests);
}
/* finished all send/recv's now poll for completion before
* continuing to next iteration
*/
completed = 0;
/* polling internally on 2*(k - 1) requests */
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
if(!completed){
/* save state and hop out
* only the iteration needs to be tracked
*/
*iteration = i; /* need to pick up here */
return ((OMPI_SUCCESS != rc) ? OMPI_ERROR : BCOL_FN_STARTED);
}
}
/* finish off the last piece, send the data back to the extra */
if( 0 < exchange_node->n_extra_sources ) {
dst = exchange_node->rank_extra_sources_array[0];
comm_dst = group_list[dst];
knt = 0;
for( i = 0; i < group_size; i++){
knt += list_connected[i];
}
/* debug print */
/*
fprintf(stderr,"sending %d bytes to extra %d \n",pack_len*knt,comm_dst);
*/
rc = MCA_PML_CALL(isend(data_buffer,
pack_len * knt,
MPI_BYTE,
comm_dst, tag,
MCA_PML_BASE_SEND_STANDARD, comm,
&(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10,("Failed to isend data"));
return OMPI_ERROR;
}
++(*active_requests);
/* probe for send completion */
completed = 0;
/* polling internally */
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
if(!completed){
/* save state and hop out
* We really do need to block here so set
* the iteration to pow_k +1 indicating we need to
* finish progressing the last part
*/
*iteration = pow_k + 1;
return (OMPI_SUCCESS != rc ? OMPI_ERROR : BCOL_FN_STARTED);
}
}
FINISHED:
/* recycle buffer if need be */
return BCOL_FN_COMPLETE;
}
/* allgather progress function */
int bcol_ptpcoll_k_nomial_allgather_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
{
/* local variables */
mca_bcol_ptpcoll_module_t *ptpcoll_module = (mca_bcol_ptpcoll_module_t *) const_args->bcol_module;
int *group_list = ptpcoll_module->super.sbgp_partner_module->group_list;
netpatterns_k_exchange_node_t *exchange_node = &ptpcoll_module->knomial_allgather_tree;
int group_size = ptpcoll_module->group_size;
int *list_connected = ptpcoll_module->super.list_n_connected; /* critical for hierarchical colls */
int tag;
int i, j;
int knt;
int comm_src, comm_dst, src, dst;
int recv_offset, recv_len;
int send_offset, send_len;
uint32_t buffer_index = input_args->buffer_index;
int pow_k, 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 = 0; /* initialized */
void *data_buffer = (void*)(
(unsigned char *) input_args->sbuf +
(size_t) input_args->sbuf_offset);
int pack_len = input_args->count * input_args->dtype->super.size;
/* initialize the counter */
int *iteration = &ptpcoll_module->ml_mem.ml_buf_desc[buffer_index].iteration;
#if 0
fprintf(stderr,"%d: entering p2p allgather progress AR: %d iter: %d\n",my_group_index,*active_requests,
*iteration);
#endif
/* keep tag within the limit supported 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;
/* k-nomial tree parameters */
tree_order = exchange_node->tree_order;
pow_k = exchange_node->log_tree_order;
/* let's begin the collective, starting with extra ranks and their
* respective proxies
*/
if( EXTRA_NODE == exchange_node->node_type ) {
/* debug print */
/*fprintf(stderr,"666 \n");*/
/* simply poll for completion */
completed = 0;
/* polling internally */
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
if(completed){
/* go to buffer release */
goto FINISHED;
}else{
/* save state and hop out
* nothing to save here
*/
return ((OMPI_SUCCESS != rc) ? OMPI_ERROR : BCOL_FN_STARTED);
}
}else if ( 0 < exchange_node->n_extra_sources && (-1 == *iteration)) {
/* I am a proxy for someone */
/* Simply poll for completion */
completed = 0;
/* polling internally */
assert( 1 == *active_requests);
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
if(!completed){
/* save state and hop out
* We really do need to block here so set
* the iteration to -1 indicating we need to
* finish this part first
*/
(*iteration) = -1;
return ((OMPI_SUCCESS != rc) ? OMPI_ERROR : BCOL_FN_STARTED);
}
/* I may now proceed to the recursive k - ing phase */
*iteration = 0;
}
/* the ordering here between the extra rank and progress active requests
* is critical
*/
/* extra rank */
if( (pow_k + 1) == *iteration ){
/* finish off the last one */
goto PROGRESS_EXTRA;
}
/* active requests must be completed before continuing on to
* recursive k -ing step
* CAREFUL HERE, IT THIS REALLY WHAT YOU WANT??
*/
if( 0 < (*active_requests) ) {
/* then we have something to progress from last step */
/* debug print */
/*
fprintf(stderr,"%d: entering progress AR: %d iter: %d\n",my_group_index,*active_requests,
*iteration);
*/
completed = 0;
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
if(!completed){
/* save state and hop out
* state hasn't changed
*/
return ((MPI_SUCCESS != rc) ? OMPI_ERROR : BCOL_FN_STARTED);
}
++(*iteration);
}
/* we start the recursive k - ing phase */
for( i = *iteration; i < pow_k; i++) {
/* nothing changes here */
for(j = 0; j < (tree_order - 1); j++) {
/* send phase */
dst = exchange_node->rank_exchanges[i][j];
if( dst < 0 ){
continue;
}
comm_dst = group_list[dst];
send_offset = exchange_node->payload_info[i][j].s_offset * pack_len;
send_len = exchange_node->payload_info[i][j].s_len * pack_len;
rc = MCA_PML_CALL(isend((void*)((unsigned char *) data_buffer +
send_offset),
send_len,
MPI_BYTE,
comm_dst, tag,
MCA_PML_BASE_SEND_STANDARD, comm,
&(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10,("Failed to isend data"));
return OMPI_ERROR;
}
++(*active_requests);
/* sends are posted */
}
/* Now post the recv's */
for( j = 0; j < (tree_order - 1); j++ ) {
/* recv phase */
src = exchange_node->rank_exchanges[i][j];
if( src < 0 ) {
continue;
}
comm_src = group_list[src];
recv_offset = exchange_node->payload_info[i][j].r_offset * pack_len;
recv_len = exchange_node->payload_info[i][j].r_len * pack_len;
/* post the receive */
rc = MCA_PML_CALL(irecv((void *) ((unsigned char *) data_buffer +
recv_offset),
recv_len,
MPI_BYTE,
comm_src,
tag, comm, &(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10, ("Failed to post ireceive "));
return OMPI_ERROR;
}
++(*active_requests);
}
/* finished all send/recv's now poll for completion before
* continuing to next iteration
*/
completed = 0;
/* make this non-blocking */
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
if(!completed){
/* save state and hop out
* We really do need to block here so set
* the iteration to -1 indicating we need to
* finish this part first
*/
*iteration = i; /* need to pick up here */
return ((OMPI_SUCCESS != rc) ? OMPI_ERROR : BCOL_FN_STARTED);
}
}
/* finish off the last piece, send the data back to the extra */
if( 0 < exchange_node->n_extra_sources ) {
dst = exchange_node->rank_extra_sources_array[0];
comm_dst = group_list[dst];
knt = 0;
for( i = 0; i < group_size; i++){
knt += list_connected[i];
}
rc = MCA_PML_CALL(isend(data_buffer,
pack_len * knt,
MPI_BYTE,
comm_dst, tag,
MCA_PML_BASE_SEND_STANDARD, comm,
&(requests[*active_requests])));
if( OMPI_SUCCESS != rc ) {
PTPCOLL_VERBOSE(10,("Failed to isend data"));
return OMPI_ERROR;
}
++(*active_requests);
/* probe for send completion */
completed = 0;
/* make this non-blocking */
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
if(!completed){
/* save state and hop out
* We really do need to block here so set
* the iteration to pow_k +1 indicating we need to
* finish progressing the last part
*/
*iteration = pow_k + 1;
return ((OMPI_SUCCESS != rc) ? OMPI_ERROR : BCOL_FN_STARTED);
}
}
/* folks need to skip this unless they really are the proxy
* reentering with the intent of progressing the final send
*/
goto FINISHED;
PROGRESS_EXTRA:
/* probe for send completion */
completed = 0;
/* make this non-blocking */
completed = mca_bcol_ptpcoll_test_all_for_match(active_requests, requests, &rc);
if(!completed){
/* save state and hop out
* We really do need to block here so set
* the iteration to pow_k +1 indicating we need to
* finish progressing the last part
*/
return ((OMPI_SUCCESS != rc) ? OMPI_ERROR : BCOL_FN_STARTED);
}
FINISHED:
/* recycle buffer if need be */
return BCOL_FN_COMPLETE;
}
/*
* Register allreduce functions to the BCOL function table,
* so they can be selected
*/
int bcol_ptpcoll_allgather_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;
comm_attribs.bcoll_type = BCOL_ALLGATHER;
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_k_nomial_allgather_init,
bcol_ptpcoll_k_nomial_allgather_progress);
comm_attribs.data_src = DATA_SRC_KNOWN;
inv_attribs.bcol_msg_min = 10000000;
inv_attribs.bcol_msg_max = 10485760; /* range 4 */
mca_bcol_base_set_attributes(super, &comm_attribs, &inv_attribs,
bcol_ptpcoll_k_nomial_allgather_init,
bcol_ptpcoll_k_nomial_allgather_progress);
return OMPI_SUCCESS;
}
Копировать постоянную ссылку