1
1
openmpi/ompi/mca/bcol/basesmuma/bcol_basesmuma_allgather.c

540 строки
19 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) 2013-2014 Los Alamos National Security, LLC. All rights
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
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "ompi/include/ompi/constants.h"
#include "ompi/mca/bcol/base/base.h"
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
#include "ompi/mca/bcol/bcol.h"
#include "ompi/mca/bcol/basesmuma/bcol_basesmuma.h"
/*
#define IS_AGDATA_READY(peer, my_flag, my_sequence_number)\
(((peer)->sequence_number == (my_sequence_number) && \
(peer)->flags[ALLGATHER_FLAG][bcol_id] >= (my_flag) \
)? true : false )
*/
#define CALC_ACTIVE_REQUESTS(active_requests,peers, tree_order) \
do{ \
for( j = 0; j < (tree_order - 1); j++){ \
if( 0 > peers[j] ) { \
/* set the bit */ \
*active_requests ^= (1<<j); \
} \
} \
}while(0)
/*
* 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_basesmuma_k_nomial_allgather_init(bcol_function_args_t *input_args,
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 */
int8_t flag_offset;
volatile int8_t ready_flag;
mca_bcol_basesmuma_module_t *bcol_module = (mca_bcol_basesmuma_module_t *) const_args->bcol_module;
netpatterns_k_exchange_node_t *exchange_node = &bcol_module->knomial_allgather_tree;
int group_size = bcol_module->colls_no_user_data.size_of_group;
int *list_connected = bcol_module->super.list_n_connected; /* critical for hierarchical colls */
int bcol_id = (int) bcol_module->super.bcol_id;
mca_bcol_basesmuma_component_t *cm = &mca_bcol_basesmuma_component;
uint32_t buffer_index = input_args->buffer_index;
int *active_requests =
&(bcol_module->ml_mem.nb_coll_desc[buffer_index].active_requests);
int *iteration = &bcol_module->ml_mem.nb_coll_desc[buffer_index].iteration;
int *status = &bcol_module->ml_mem.nb_coll_desc[buffer_index].status;
int leading_dim, buff_idx, idx;
int i, j, probe;
int knt;
int src;
int recv_offset, recv_len;
int pow_k, tree_order;
int max_requests = 0; /* important to initialize this */
int matched = 0;
int64_t sequence_number=input_args->sequence_num;
int my_rank = bcol_module->super.sbgp_partner_module->my_index;
int buff_offset = bcol_module->super.hier_scather_offset;
int pack_len = input_args->count * input_args->dtype->super.size;
void *data_addr = (void*)(
(unsigned char *) input_args->sbuf +
(size_t) input_args->sbuf_offset);
volatile mca_bcol_basesmuma_payload_t *data_buffs;
volatile char *peer_data_pointer;
/* control structures */
volatile mca_bcol_basesmuma_header_t *my_ctl_pointer;
volatile mca_bcol_basesmuma_header_t *peer_ctl_pointer;
#if 0
fprintf(stderr,"entering p2p allgather pack_len %d\n",pack_len);
#endif
/* initialize the iteration counter */
buff_idx = input_args->src_desc->buffer_index;
leading_dim = bcol_module->colls_no_user_data.size_of_group;
idx=SM_ARRAY_INDEX(leading_dim,buff_idx,0);
data_buffs=(volatile mca_bcol_basesmuma_payload_t *)
bcol_module->colls_with_user_data.data_buffs+idx;
/* Set pointer to current proc ctrl region */
my_ctl_pointer = data_buffs[my_rank].ctl_struct;
/* NTH: copied from progress */
flag_offset = my_ctl_pointer->starting_flag_value[bcol_id];
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
/* initialize headers and ready flag */
BASESMUMA_HEADER_INIT(my_ctl_pointer, ready_flag, sequence_number, bcol_id);
/* initialize these */
*iteration = 0;
*active_requests = 0;
*status = 0;
/* k-nomial parameters */
tree_order = exchange_node->tree_order;
pow_k = exchange_node->log_tree_order;
/* calculate the maximum number of requests
* at each level each rank communicates with
* at most (k - 1) peers
* so if we set k - 1 bit fields in "max_requests", then
* we have max_request == 2^(k - 1) -1
*/
for(i = 0; i < (tree_order - 1); i++){
max_requests ^= (1<<i);
}
/* let's begin the collective, starting with extra ranks and their
* respective proxies
*/
if( EXTRA_NODE == exchange_node->node_type ) {
/* then I will signal to my proxy rank*/
my_ctl_pointer->flags[ALLGATHER_FLAG][bcol_id] = ready_flag;
ready_flag = flag_offset + 1 + pow_k + 2;
/* now, poll for completion */
src = exchange_node->rank_extra_sources_array[0];
peer_data_pointer = data_buffs[src].payload;
peer_ctl_pointer = data_buffs[src].ctl_struct;
/* calculate the offset */
knt = 0;
for(i = 0; i < group_size; i++){
knt += list_connected[i];
}
for( i = 0; i < cm->num_to_probe && (0 == matched); i++ ) {
if(IS_PEER_READY(peer_ctl_pointer, ready_flag, sequence_number, ALLGATHER_FLAG, bcol_id)){
matched = 1;
/* we receive the entire message */
memcpy((void *)((unsigned char *) data_addr + buff_offset),
(void *) ((unsigned char *) peer_data_pointer + buff_offset),
knt * pack_len);
goto FINISHED;
}
}
/* save state and bail */
*iteration = -1;
return 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];
peer_data_pointer = data_buffs[src].payload;
peer_ctl_pointer = data_buffs[src].ctl_struct;
knt = 0;
for(i = 0; i < src; i++){
knt += list_connected[i];
}
/* probe for extra rank's arrival */
for( i = 0; i < cm->num_to_probe && ( 0 == matched); i++) {
if(IS_PEER_READY(peer_ctl_pointer,ready_flag, sequence_number, ALLGATHER_FLAG, bcol_id)){
matched = 1;
/* copy it in */
memcpy((void *)((unsigned char *) data_addr + knt*pack_len),
(void *) ((unsigned char *) peer_data_pointer + knt*pack_len),
pack_len * list_connected[src]);
goto MAIN_PHASE;
}
}
*status = ready_flag;
*iteration = -1;
return BCOL_FN_STARTED;
}
MAIN_PHASE:
/* bump the ready flag */
ready_flag++;
/* we start the recursive k - ing phase */
for( *iteration = 0; *iteration < pow_k; (*iteration)++) {
/* announce my arrival */
opal_atomic_wmb ();
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
my_ctl_pointer->flags[ALLGATHER_FLAG][bcol_id] = ready_flag;
/* calculate the number of active requests */
CALC_ACTIVE_REQUESTS(active_requests,exchange_node->rank_exchanges[*iteration],tree_order);
/* Now post the recv's */
for( j = 0; j < (tree_order - 1); j++ ) {
/* recv phase */
src = exchange_node->rank_exchanges[*iteration][j];
if( src < 0 ) {
/* then not a valid rank, continue */
continue;
}
peer_data_pointer = data_buffs[src].payload;
peer_ctl_pointer = data_buffs[src].ctl_struct;
if( !(*active_requests&(1<<j))) {
/* then the bit hasn't been set, thus this peer
* hasn't been processed at this level
*/
recv_offset = exchange_node->payload_info[*iteration][j].r_offset * pack_len;
recv_len = exchange_node->payload_info[*iteration][j].r_len * pack_len;
/* post the receive */
/* I am putting the probe loop as the inner most loop to achieve
* better temporal locality
*/
matched = 0;
for( probe = 0; probe < cm->num_to_probe && (0 == matched); probe++){
if(IS_PEER_READY(peer_ctl_pointer,ready_flag, sequence_number, ALLGATHER_FLAG, bcol_id)){
matched = 1;
/* set this request's bit */
*active_requests ^= (1<<j);
/* get the data */
memcpy((void *)((unsigned char *) data_addr + recv_offset),
(void *)((unsigned char *) peer_data_pointer + recv_offset),
recv_len);
}
}
}
}
if( max_requests == *active_requests ){
/* bump the ready flag */
ready_flag++;
/*reset the active requests */
*active_requests = 0;
} else {
/* save state and hop out
* only the iteration needs to be tracked
*/
*status = my_ctl_pointer->flags[ALLGATHER_FLAG][bcol_id];
return BCOL_FN_STARTED;
}
}
/* bump the flag one more time for the extra rank */
ready_flag = flag_offset + 1 + pow_k + 2;
/* finish off the last piece, send the data back to the extra */
if( 0 < exchange_node->n_extra_sources ) {
/* simply announce my arrival */
opal_atomic_wmb ();
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
my_ctl_pointer->flags[ALLGATHER_FLAG][bcol_id] = ready_flag;
}
FINISHED:
/* bump this up */
my_ctl_pointer->starting_flag_value[bcol_id]++;
return BCOL_FN_COMPLETE;
}
/* allgather progress function */
int bcol_basesmuma_k_nomial_allgather_progress(bcol_function_args_t *input_args,
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 */
int8_t flag_offset;
uint32_t buffer_index = input_args->buffer_index;
volatile int8_t ready_flag;
mca_bcol_basesmuma_module_t *bcol_module = (mca_bcol_basesmuma_module_t *) const_args->bcol_module;
netpatterns_k_exchange_node_t *exchange_node = &bcol_module->knomial_allgather_tree;
int group_size = bcol_module->colls_no_user_data.size_of_group;
int *list_connected = bcol_module->super.list_n_connected; /* critical for hierarchical colls */
int bcol_id = (int) bcol_module->super.bcol_id;
mca_bcol_basesmuma_component_t *cm = &mca_bcol_basesmuma_component;
int *active_requests =
&(bcol_module->ml_mem.nb_coll_desc[buffer_index].active_requests);
int *iteration = &bcol_module->ml_mem.nb_coll_desc[buffer_index].iteration;
int *iter = iteration; /* double alias */
int *status = &bcol_module->ml_mem.nb_coll_desc[buffer_index].status;
int leading_dim, idx, buff_idx;
int i, j, probe;
int knt;
int src;
int recv_offset, recv_len;
int max_requests = 0; /* critical to set this */
int pow_k, tree_order;
int matched = 0;
int64_t sequence_number=input_args->sequence_num;
int my_rank = bcol_module->super.sbgp_partner_module->my_index;
int buff_offset = bcol_module->super.hier_scather_offset;
int pack_len = input_args->count * input_args->dtype->super.size;
void *data_addr = (void*)(
(unsigned char *) input_args->sbuf +
(size_t) input_args->sbuf_offset);
volatile mca_bcol_basesmuma_payload_t *data_buffs;
volatile char *peer_data_pointer;
/* control structures */
volatile mca_bcol_basesmuma_header_t *my_ctl_pointer;
volatile mca_bcol_basesmuma_header_t *peer_ctl_pointer;
#if 0
fprintf(stderr,"%d: entering sm allgather progress active requests %d iter %d ready_flag %d\n",my_rank,
*active_requests,*iter,*status);
#endif
buff_idx = input_args->src_desc->buffer_index;
leading_dim=bcol_module->colls_no_user_data.size_of_group;
idx=SM_ARRAY_INDEX(leading_dim,buff_idx,0);
data_buffs=(volatile mca_bcol_basesmuma_payload_t *)
bcol_module->colls_with_user_data.data_buffs+idx;
/* Set pointer to current proc ctrl region */
my_ctl_pointer = data_buffs[my_rank].ctl_struct;
/* increment the starting flag by one and return */
/* flag offset seems unnecessary here */
flag_offset = my_ctl_pointer->starting_flag_value[bcol_id];
ready_flag = *status;
my_ctl_pointer->sequence_number = sequence_number;
/* k-nomial parameters */
tree_order = exchange_node->tree_order;
pow_k = exchange_node->log_tree_order;
/* calculate the maximum number of requests
* at each level each rank communicates with
* at most (k - 1) peers
* so if we set k - 1 bit fields in "max_requests", then
* we have max_request == 2^(k - 1) -1
*/
for(i = 0; i < (tree_order - 1); i++){
max_requests ^= (1<<i);
}
/* let's begin the collective, starting with extra ranks and their
* respective proxies
*/
if( EXTRA_NODE == exchange_node->node_type ) {
/* If I'm in here, then I must be looking for data */
ready_flag = flag_offset + 1 + pow_k + 2;
src = exchange_node->rank_extra_sources_array[0];
peer_data_pointer = data_buffs[src].payload;
peer_ctl_pointer = data_buffs[src].ctl_struct;
/* calculate the offset */
knt = 0;
for(i = 0; i < group_size; i++){
knt += list_connected[i];
}
for( i = 0; i < cm->num_to_probe && (0 == matched); i++ ) {
if(IS_PEER_READY(peer_ctl_pointer, ready_flag, sequence_number, ALLGATHER_FLAG, bcol_id)){
matched = 1;
/* we receive the entire message */
memcpy((void *)((unsigned char *) data_addr + buff_offset),
(void *) ((unsigned char *) peer_data_pointer + buff_offset),
knt * pack_len);
goto FINISHED;
}
}
/* haven't found it, state is saved, bail out */
return BCOL_FN_STARTED;
}else if ( ( -1 == *iteration ) && (0 < exchange_node->n_extra_sources) ) {
/* I am a proxy for someone */
src = exchange_node->rank_extra_sources_array[0];
peer_data_pointer = data_buffs[src].payload;
peer_ctl_pointer = data_buffs[src].ctl_struct;
knt = 0;
for(i = 0; i < src; i++){
knt += list_connected[i];
}
/* probe for extra rank's arrival */
for( i = 0; i < cm->num_to_probe && ( 0 == matched); i++) {
if(IS_PEER_READY(peer_ctl_pointer,ready_flag, sequence_number, ALLGATHER_FLAG, bcol_id)){
matched = 1;
/* copy it in */
memcpy((void *)((unsigned char *) data_addr + knt*pack_len),
(void *) ((unsigned char *) peer_data_pointer + knt*pack_len),
pack_len * list_connected[src]);
ready_flag++;
*iteration = 0;
goto MAIN_PHASE;
}
}
return BCOL_FN_STARTED;
}
MAIN_PHASE:
/* start the recursive k - ing phase */
for( *iter=*iteration; *iter < pow_k; (*iter)++) {
/* I am ready at this level */
opal_atomic_wmb ();
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
my_ctl_pointer->flags[ALLGATHER_FLAG][bcol_id] = ready_flag;
if( 0 == *active_requests ) {
/* flip some bits, if we don't have active requests from a previous visit */
CALC_ACTIVE_REQUESTS(active_requests,exchange_node->rank_exchanges[*iter],tree_order);
}
for( j = 0; j < (tree_order - 1); j++ ) {
/* recv phase */
src = exchange_node->rank_exchanges[*iter][j];
if( src < 0 ) {
/* then not a valid rank, continue
*/
continue;
}
peer_data_pointer = data_buffs[src].payload;
peer_ctl_pointer = data_buffs[src].ctl_struct;
if( !(*active_requests&(1<<j))){
/* then this peer hasn't been processed at this level */
recv_offset = exchange_node->payload_info[*iter][j].r_offset * pack_len;
recv_len = exchange_node->payload_info[*iter][j].r_len * pack_len;
/* I am putting the probe loop as the inner most loop to achieve
* better temporal locality
*/
matched = 0;
for( probe = 0; probe < cm->num_to_probe && (0 == matched); probe++){
if(IS_PEER_READY(peer_ctl_pointer,ready_flag, sequence_number, ALLGATHER_FLAG, bcol_id)){
matched = 1;
/* flip the request's bit */
*active_requests ^= (1<<j);
/* copy the data */
memcpy((void *)((unsigned char *) data_addr + recv_offset),
(void *)((unsigned char *) peer_data_pointer + recv_offset),
recv_len);
}
}
}
}
if( max_requests == *active_requests ){
/* bump the ready flag */
ready_flag++;
/* reset the active requests for the next level */
*active_requests = 0;
/* calculate the number of active requests
* logically makes sense to do it here. We don't
* want to inadvertantly flip a bit to zero that we
* set previously
*/
} else {
/* state is saved hop out
*/
*status = my_ctl_pointer->flags[ALLGATHER_FLAG][bcol_id];
return BCOL_FN_STARTED;
}
}
/* bump the flag one more time for the extra rank */
ready_flag = flag_offset + 1 + pow_k + 2;
/* finish off the last piece, send the data back to the extra */
if( 0 < exchange_node->n_extra_sources ) {
/* simply announce my arrival */
opal_atomic_wmb ();
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
my_ctl_pointer->flags[ALLGATHER_FLAG][bcol_id] = ready_flag;
}
FINISHED:
/* bump this up for others to see */
my_ctl_pointer->starting_flag_value[bcol_id]++;
return BCOL_FN_COMPLETE;
}
/* Register allreduce functions to the BCOL function table,
* so they can be selected
*/
int bcol_basesmuma_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_basesmuma_k_nomial_allgather_init,
bcol_basesmuma_k_nomial_allgather_progress);
return OMPI_SUCCESS;
}