ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
22f64bb62b
openmpi/ompi/mca/bcol/basesmuma/bcol_basesmuma_gather.c

1107 строки
40 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) 2009-2012 Oak Ridge National Laboratory. All rights reserved.
* Copyright (c) 2009-2012 Mellanox Technologies. All rights reserved.
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
* 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"
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
#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/basesmuma/bcol_basesmuma.h"
#include "ompi/constants.h"
#include "ompi/datatype/ompi_datatype.h"
#include "ompi/communicator/communicator.h"
/* debug
* #include "opal/sys/timer.h"
*
* extern uint64_t timers[7];
* end debug */
/* debug */
#include <unistd.h>
/* end debug */
/* non-blocking gather routines: init and progress functions */
int bcol_basesmuma_gather_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_GATHER;
comm_attribs.comm_size_min = 0;
Addendum to r31096. Up basesmuma algorithm limits to 1M. After discussion with Manju we decided to update these the process count limits of the shared memory collectives to an arbitrarily large number. cmr=v1.7.5:ticket=trac:4405 This commit was SVN r31126. The following SVN revision numbers were found above: r31096 --> open-mpi/ompi@3f469d08e7f8930aab44e402c90618788e44d1f8 The following Trac tickets were found above: Ticket 4405 --> https://svn.open-mpi.org/trac/ompi/ticket/4405
2014-03-19 01:25:49 +04:00
comm_attribs.comm_size_max = 1048576;
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
comm_attribs.data_src = DATA_SRC_KNOWN;
comm_attribs.waiting_semantics = BLOCKING;
inv_attribs.bcol_msg_min = 0;
inv_attribs.bcol_msg_max = 20000;
inv_attribs.datatype_bitmap = 0x11111111;
inv_attribs.op_types_bitmap = 0x11111111;
/* Set attributes for fanin fanout algorithm */
mca_bcol_base_set_attributes(super, &comm_attribs, &inv_attribs,
bcol_basesmuma_k_nomial_gather_init,
bcol_basesmuma_k_nomial_gather_progress);
return OMPI_SUCCESS;
}
int bcol_basesmuma_k_nomial_gather_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
mca_bcol_base_function_t *c_input_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 */
int leading_dim, buff_idx, idx;
int src, i, j, k_temp1, k_temp2;
int pseudo_root, proxy_root, pseudo_base_adj;
volatile int8_t ready_flag;
int count=input_args->count;
struct ompi_datatype_t* dtype=input_args->dtype;
int root=input_args->root;
int base_adj, base;
int total_peers, my_pow_k=0;
int64_t sequence_number=input_args->sequence_num;
mca_bcol_basesmuma_module_t* bcol_module=
(mca_bcol_basesmuma_module_t *)c_input_args->bcol_module;
int bcol_id = (int) bcol_module->super.bcol_id;
int my_rank = bcol_module->super.sbgp_partner_module->my_index;
netpatterns_k_exchange_node_t *exchange_node =
&bcol_module->knomial_allgather_tree;
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 buff_offset = bcol_module->super.hier_scather_offset;
/* "indirectors" */
int *inv_map = exchange_node->inv_reindex_map;
int *reindex_map = exchange_node->reindex_map;
int stray = exchange_node->k_nomial_stray;
/* tree radix */
int tree_order = exchange_node->tree_order;
/* tree depth */
int pow_k = exchange_node->log_tree_order;
/* largest power of k less than or equal to np */
int cnt = exchange_node->n_largest_pow_tree_order;
/* payload structures */
volatile mca_bcol_basesmuma_payload_t *data_buffs;
/* control structures */
volatile mca_bcol_basesmuma_header_t *my_ctl_pointer;
size_t pack_len = 0, dt_size;
#if 0
fprintf(stderr,"Entering sm gather input_args->sbuf_offset %d \n",input_args->sbuf_offset);
fflush(stderr);
#endif
/* we will work only on packed data - so compute the length*/
/* this is the size of my data, this is not gatherv so it's the same
* for all ranks in the communicator.
*/
ompi_datatype_type_size(dtype, &dt_size);
pack_len=count*dt_size;
/* now set the "real" offset */
buff_offset = buff_offset*pack_len;
buff_idx = input_args->src_desc->buffer_index;
/* Get addressing information */
my_rank = bcol_module->super.sbgp_partner_module->my_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;
/* init the header */
BASESMUMA_HEADER_INIT(my_ctl_pointer, ready_flag, sequence_number, bcol_id);
/* init active requests, iteration, and status */
*iteration = 0;
*active_requests = 0;
*status = -1;
/* calculate the number of steps necessary for this collective */
/* first thing we do is figure out where the root is in our new indexing */
/* find root in new indexing */
pseudo_root = inv_map[root];
/* see if this is larger than the stray */
if (pseudo_root >= stray) {
/* then we need to define the proxy root, everyone can do this */
proxy_root = pseudo_root - cnt;
} else {
proxy_root = pseudo_root;
}
/* do some figuring */
if (EXCHANGE_NODE == exchange_node->node_type) {
total_peers = 0;
my_pow_k = pow_k;
k_temp1 = tree_order;
k_temp2 = 1;
for( i = 0; i < pow_k; i++) {
/* then find the base */
FIND_BASE(base,exchange_node->reindex_myid,i+1,tree_order);
/* now find the adjusted base */
base_adj = base + (base + proxy_root)%k_temp1;
/* ok, now find out WHO is occupying this slot */
pseudo_base_adj = reindex_map[base_adj];
if(my_rank == pseudo_base_adj ) {
/* then go ahead and poll for children's data */
for( j = 0; j < (tree_order - 1); j++ ) {
/* send phase
*/
/* get communication partner */
src = exchange_node->rank_exchanges[i][j];
/* remember, if we have extra ranks, then we won't participate
* with a least one peer. Make a check
*/
if( src < 0 ){
continue;
}else{
/* flip a bit to represent this request */
*active_requests ^= (1<<total_peers++);
}
}
} else {
/* announce my arrival */
my_pow_k = i;
break;
}
k_temp1 = k_temp1*tree_order;
k_temp2 = k_temp2*tree_order;
}
}
*iteration = my_pow_k;
if (EXTRA_NODE == exchange_node->node_type || 0 == exchange_node->n_extra_sources) {
if (0 == my_pow_k || EXTRA_NODE == exchange_node->node_type) {
opal_atomic_rmb ();
my_ctl_pointer->flags[GATHER_FLAG][bcol_id] = ready_flag;
}
if ((EXTRA_NODE == exchange_node->node_type && root != my_rank) || 0 == my_pow_k) {
/* nothing more to do */
my_ctl_pointer->starting_flag_value[bcol_id]++;
return BCOL_FN_COMPLETE;
}
}
return BCOL_FN_STARTED;
}
int bcol_basesmuma_k_nomial_gather_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
mca_bcol_base_function_t *c_input_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 */
int group_size;
int flag_offset;
int leading_dim, buff_idx, idx;
int src, knt, i, j, k_temp1, k_temp2;
volatile int8_t ready_flag;
int count=input_args->count;
struct ompi_datatype_t* dtype=input_args->dtype;
int root=input_args->root;
int probe;
int matched;
int64_t sequence_number=input_args->sequence_num;
mca_bcol_basesmuma_module_t* bcol_module=
(mca_bcol_basesmuma_module_t *)c_input_args->bcol_module;
int bcol_id = (int) bcol_module->super.bcol_id;
int my_rank = bcol_module->super.sbgp_partner_module->my_index;
mca_bcol_basesmuma_component_t *cm = &mca_bcol_basesmuma_component;
netpatterns_k_exchange_node_t *exchange_node =
&bcol_module->knomial_allgather_tree;
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 buff_offset = bcol_module->super.hier_scather_offset;
/* "indirectors" */
int *list_connected = bcol_module->super.list_n_connected;
/* tree radix */
int tree_order = exchange_node->tree_order;
/* payload structures */
volatile mca_bcol_basesmuma_payload_t *data_buffs;
volatile char *child_data_pointer;
/* control structures */
volatile mca_bcol_basesmuma_header_t *my_ctl_pointer;
volatile mca_bcol_basesmuma_header_t *child_ctl_pointer;
/*volatile mca_bcol_basesmuma_ctl_struct_t* parent_ctl_pointer; */
size_t pack_len = 0, dt_size;
void *data_addr = (void *)((unsigned char *)input_args->src_desc->data_addr);
#if 0
fprintf(stderr,"Entering sm gather input_args->sbuf_offset %d \n",input_args->sbuf_offset);
fflush(stderr);
#endif
/* we will work only on packed data - so compute the length*/
/* this is the size of my data, this is not gatherv so it's the same
* for all ranks in the communicator.
*/
ompi_datatype_type_size(dtype, &dt_size);
pack_len=count*dt_size;
/* now set the "real" offset */
buff_offset = buff_offset*pack_len;
buff_idx = input_args->src_desc->buffer_index;
/* Get addressing information */
my_rank = bcol_module->super.sbgp_partner_module->my_index;
group_size = bcol_module->colls_no_user_data.size_of_group;
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;
/* restart the ready_flag state */
flag_offset = my_ctl_pointer->starting_flag_value[bcol_id];
ready_flag = flag_offset + 1;
/* calculate the number of steps necessary for this collective */
/* first thing we do is figure out where the root is in our new indexing */
/* find root in new indexing */
if( EXTRA_NODE == exchange_node->node_type ) {
/* poll for data from proxy */
src = exchange_node->rank_extra_sources_array[0];
/* get src data buffer */
child_data_pointer = data_buffs[src].payload;
child_ctl_pointer = data_buffs[src].ctl_struct;
/* remember to bump your flag */
ready_flag++;
/* in this case, you must block */
for (i = 0 ; i < cm->num_to_probe ; ++i) {
if (IS_PEER_READY(child_ctl_pointer,ready_flag,sequence_number, GATHER_FLAG, bcol_id)){
/* receive the data from the proxy, aka pseudo-root */
memcpy((void *) ((unsigned char *) data_addr + buff_offset),
(void *) ((unsigned char *) child_data_pointer+buff_offset),
pack_len * group_size);
goto FINISHED;
}
}
return BCOL_FN_STARTED;
}
if (0 < exchange_node->n_extra_sources && (-1 == (*status))) {
/* am a proxy, poll for pack_len data from extra */
src = exchange_node->rank_extra_sources_array[0];
/* get src data buffer */
child_data_pointer = data_buffs[src].payload;
child_ctl_pointer = data_buffs[src].ctl_struct;
knt = 0;
for( i = 0; i < src; i++){
knt += list_connected[i];
}
/* must block here also */
matched = 0;
for (i = 0, matched = 0 ; i < cm->num_to_probe && (0 == matched) ; ++i) {
if(IS_PEER_READY(child_ctl_pointer,ready_flag,sequence_number, GATHER_FLAG, bcol_id)){
matched = 1;
memcpy((void *) ((unsigned char *) data_addr + buff_offset + pack_len*knt),
(void *) ((unsigned char *) child_data_pointer + buff_offset +
pack_len*knt), pack_len*list_connected[src]);
*status = 0;
if( 0 == *active_requests ){
goto LAST_STEP;
}
break;
}
}
if( 0 == matched ){
return BCOL_FN_STARTED;
}
}
/* start the k-nomial gather phase */
/* only "active ranks participate, once a rank has forwarded its data, it becomes inactive */
for (probe = 0 ; probe < cm->num_to_probe ; ++probe) {
k_temp1 = tree_order;
k_temp2 = 1;
for (i = 0 ; i < *(iteration) ; ++i) {
/* then go ahead and poll for children's data */
for (j = 0 ; j < (tree_order - 1) ; ++j) {
/* send phase
*/
/* get communication partner */
src = exchange_node->rank_exchanges[i][j];
/* remember, if we have extra ranks, then we won't participate
* with a least one peer. Make a check
*/
/* if the bit that corresponds to this child has been set to zero,
* then it has already checked in and data received
*/
if (src < 0 || 1 != ((*active_requests >> ((tree_order - 1)*i + j))&1)){
continue;
}
child_data_pointer = data_buffs[src].payload;
child_ctl_pointer = data_buffs[src].ctl_struct;
if(IS_PEER_READY(child_ctl_pointer,ready_flag,sequence_number, GATHER_FLAG, bcol_id)){
/* copy the data */
memcpy((void *) ((unsigned char *) data_addr + buff_offset +
exchange_node->payload_info[i][j].r_offset*pack_len),
(void *) ((unsigned char *) child_data_pointer + buff_offset +
exchange_node->payload_info[i][j].r_offset*pack_len),
exchange_node->payload_info[i][j].r_len*pack_len);
/* flip the bit to zero */
*active_requests ^= (1<<((tree_order - 1)*i + j));
if(0 == (*active_requests)) {
goto LAST_STEP;
}
}
}
}
k_temp1 = k_temp1*tree_order;
k_temp2 = k_temp2*tree_order;
}
return BCOL_FN_STARTED;
LAST_STEP:
/* last step, proxies send full data back to the extra ranks */
if( 0 < exchange_node->n_extra_sources &&
root == exchange_node->rank_extra_sources_array[0]) {
/* regardless, I will bump the ready flag and set it in case someone is watching */
/* announce that data is ready */
ready_flag++;
}
/* signal that data is ready */
opal_atomic_wmb ();
my_ctl_pointer->flags[GATHER_FLAG][bcol_id] = ready_flag;
FINISHED:
my_ctl_pointer->starting_flag_value[bcol_id]++;
return BCOL_FN_COMPLETE;
}
/* Blocking routines, used to prototype and test signaling,
* as well as debug hierarchical algorithm
*/
#if 0
int bcol_basesmuma_gather_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_GATHER;
comm_attribs.comm_size_min = 0;
comm_attribs.comm_size_max = 16;
comm_attribs.data_src = DATA_SRC_KNOWN;
comm_attribs.waiting_semantics = BLOCKING;
inv_attribs.bcol_msg_min = 0;
inv_attribs.bcol_msg_max = 20000;
inv_attribs.datatype_bitmap = 0x11111111;
inv_attribs.op_types_bitmap = 0x11111111;
/* Set attributes for fanin fanout algorithm */
mca_bcol_base_set_attributes(super, &comm_attribs, &inv_attribs, bcol_basesmuma_k_nomial_gather,
bcol_basesmuma_k_nomial_gather);
return OMPI_SUCCESS;
}
#endif
/* original, fully blocking, fully synchronous gather - should result in worst performance when used */
#if 0
int bcol_basesmuma_k_nomial_gather(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
mca_bcol_base_function_t *c_input_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 */
int group_size;
int first_instance=0, flag_offset;
int rc = OMPI_SUCCESS;
int leading_dim, buff_idx, idx;
int *group_list;
int src, comm_src, knt, i, k, j, k_temp1, k_temp2;
int pseudo_root, proxy_root, pseudo_base_adj;
volatile int64_t ready_flag;
int count=input_args->count;
struct ompi_datatype_t* dtype=input_args->dtype;
int root=input_args->root;
int base_adj, base;
int64_t sequence_number=input_args->sequence_num;
mca_bcol_basesmuma_module_t* bcol_module=
(mca_bcol_basesmuma_module_t *)c_input_args->bcol_module;
int my_rank = bcol_module->super.sbgp_partner_module->my_index;
mca_bcol_basesmuma_component_t *cs = &mca_bcol_basesmuma_component;
netpatterns_k_exchange_node_t *exchange_node =
&bcol_module->knomial_allgather_tree;
int buff_offset = bcol_module->super.hier_scather_offset;
/* "indirectors" */
int *list_connected = bcol_module->super.list_n_connected;
int *inv_map = exchange_node->inv_reindex_map;
int *reindex_map = exchange_node->reindex_map;
/*int *reindex_map = exchange_node->reindex_map;*/
/* stray rank == first rank in the extra set */
int stray = exchange_node->k_nomial_stray;
/* tree radix */
int tree_order = exchange_node->tree_order;
/* tree depth */
int pow_k = exchange_node->log_tree_order;
/* largest power of k less than or equal to np */
int cnt = exchange_node->n_largest_pow_tree_order;
/*fprintf(stderr,"tree order %d pow_k %d stray %d root %d\n",tree_order, pow_k, stray, root);*/
/* payload structures */
volatile mca_bcol_basesmuma_payload_t *data_buffs;
volatile char *child_data_pointer;
/* control structures */
volatile mca_bcol_basesmuma_header_t *my_ctl_pointer;
volatile mca_bcol_basesmuma_header_t *child_ctl_pointer;
/*volatile mca_bcol_basesmuma_ctl_struct_t* parent_ctl_pointer; */
size_t pack_len = 0, dt_size;
void *data_addr = (void *)((unsigned char *)input_args->src_desc->data_addr);
/* active in the algorithm */
bool active = true;
#if 0
fprintf(stderr,"Entering sm gather input_args->sbuf_offset %d \n",input_args->sbuf_offset);
fflush(stderr);
#endif
/* we will work only on packed data - so compute the length*/
/* this is the size of my data, this is not gatherv so it's the same
* for all ranks in the communicator.
*/
ompi_datatype_type_size(dtype, &dt_size);
pack_len=count*dt_size;
/* now set the "real" offset */
buff_offset = buff_offset*pack_len;
buff_idx = input_args->src_desc->buffer_index;
/* Get addressing information */
my_rank = bcol_module->super.sbgp_partner_module->my_index;
/* I have a feeling that I'll need this */
group_list = bcol_module->super.sbgp_partner_module->group_list;
group_size = bcol_module->colls_no_user_data.size_of_group;
leading_dim=bcol_module->colls_no_user_data.size_of_group;
idx=SM_ARRAY_INDEX(leading_dim,buff_idx,0);
/*ctl_structs=(mca_bcol_basesmuma_ctl_struct_t **)
bcol_module->colls_with_user_data.ctl_buffs+idx;
*/
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 = ctl_structs[my_rank]; */
my_ctl_pointer = data_buffs[my_rank].ctl_struct;
/* setup resource recycling */
if( my_ctl_pointer->sequence_number < sequence_number ) {
first_instance=1;
}
if( first_instance ) {
/* Signal arrival */
my_ctl_pointer->flag = -1;
my_ctl_pointer->gflag = -1;
my_ctl_pointer->index=1;
/* this does not need to use any flag values , so only need to
* set the value for subsequent values that may need this */
my_ctl_pointer->starting_flag_value=0;
flag_offset=0;
} else {
/* only one thread at a time will be making progress on this
* collective, so no need to make this atomic */
my_ctl_pointer->index++;
}
/* increment the starting flag by one and return */
flag_offset = my_ctl_pointer->starting_flag_value;
ready_flag = flag_offset + sequence_number + 1;
my_ctl_pointer->sequence_number = sequence_number;
/* debug
fprintf(stderr," sequence_number %lld flag_offset %d starting flag val %d\n",sequence_number,flag_offset, my_ctl_pointer->starting_flag_value);
fflush(stderr);
end debug */
/*
* Fan out from root
*/
/* don't need this either */
/* root is the local leader */
/* calculate the number of steps necessary for this collective */
/* first thing we do is figure out where the root is in our new indexing */
/* find root in new indexing */
pseudo_root = inv_map[root];
/* see if this is larger than the stray */
if( pseudo_root >= stray ) {
/* then we need to define the proxy root, everyone can do this */
proxy_root = pseudo_root - cnt;
}else {
proxy_root = pseudo_root;
}
if( EXTRA_NODE == exchange_node->node_type ) {
/* signal arrival */
my_ctl_pointer->gflag = ready_flag;
/* send is done */
/* poll for data only if I am the root */
/* bump the ready flag */
ready_flag++;
if( root == my_rank ){
/* poll for data from proxy */
src = exchange_node->rank_extra_sources_array[0];
/* get src data buffer */
child_data_pointer = data_buffs[src].payload;
child_ctl_pointer = data_buffs[src].ctl_struct;
while(!IS_GDATA_READY(child_ctl_pointer,ready_flag,sequence_number)){
opal_progress();
}
/* receive the data from the proxy, aka pseudo-root */
memcpy((void *) ((unsigned char *) data_addr + buff_offset),(void *) ((unsigned char *) child_data_pointer+buff_offset)
,pack_len*group_size);
}
goto FINISHED;
} else if( 0 < exchange_node->n_extra_sources ) {
/* am a proxy, poll for pack_len data from extra */
src = exchange_node->rank_extra_sources_array[0];
/* get src data buffer */
child_data_pointer = data_buffs[src].payload;
child_ctl_pointer = data_buffs[src].ctl_struct;
knt = 0;
for( i = 0; i < src; i++){
knt += list_connected[i];
}
while(!IS_GDATA_READY(child_ctl_pointer,ready_flag,sequence_number)){
opal_progress();
}
memcpy((void *) ((unsigned char *) data_addr + buff_offset + pack_len*knt),
(void *) ((unsigned char *) child_data_pointer + buff_offset +
pack_len*knt), pack_len*list_connected[src]);
/*fprintf(stderr,"999 proxy received data from %d at offset %d of length %d\n",src,
buff_offset+pack_len*knt,pack_len*list_connected[src]);
*/
}
/* start the k-nomial gather phase */
/* only "active ranks participate, once a rank has forwarded its data, it becomes inactive */
knt = 0;
while(active){
k_temp1 = tree_order;
k_temp2 = 1;
for( i = 0; i < pow_k; i++) {
/* then find the base */
/*FIND_BASE(base,my_rank,i+1,tree_order);*/
FIND_BASE(base,exchange_node->reindex_myid,i+1,tree_order);
/* now find the adjusted base */
base_adj = base + (base + proxy_root)%k_temp1;
/* ok, now find out WHO is occupying this slot */
/*pseudo_base_adj = inv_map[base_adj];*/
pseudo_base_adj = reindex_map[base_adj];
if(my_rank == pseudo_base_adj ) {
/* then go ahead and poll for children's data */
for( j = 0; j < (tree_order - 1); j++ ) {
/* send phase
*/
/* get communication partner */
src = exchange_node->rank_exchanges[i][j];
/*fprintf(stderr,"comm_src %d\n",comm_src);*/
/* remember, if we have extra ranks, then we won't participate
* with a least one peer. Make a check
*/
if( src < 0 ){
continue;
}
/*fprintf(stderr,"src %d\n",src);*/
child_data_pointer = data_buffs[src].payload;
child_ctl_pointer = data_buffs[src].ctl_struct;
while(!IS_GDATA_READY(child_ctl_pointer,ready_flag,sequence_number)){
opal_progress();
}
memcpy((void *) ((unsigned char *) data_addr + buff_offset +
exchange_node->payload_info[i][j].r_offset*pack_len),
(void *) ((unsigned char *) child_data_pointer + buff_offset +
exchange_node->payload_info[i][j].r_offset*pack_len),
exchange_node->payload_info[i][j].r_len*pack_len);
/*
fprintf(stderr,"999 receiving data from %d at offset %d of length %d\n",
exchange_node->rank_exchanges[i][j], buff_offset + exchange_node->payload_info[i][j].r_offset,
exchange_node->payload_info[i][j].r_len*pack_len);
*/
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
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
knt++;
if(knt == exchange_node->n_actual_exchanges) {
/* this is the trick to break the root out,
* only the root should be able to satisfy this
*/
/*
fprintf(stderr,"hello n_actual is %d \n",knt);
fprintf(stderr,"hello n_actual_exch is %d \n",
exchange_node->n_actual_exchanges);
*/
goto LAST_STEP;
}
}
} else {
/* announce my arrival */
my_ctl_pointer->gflag = ready_flag;
active = false;
break;
}
k_temp1 = k_temp1*tree_order;
k_temp2 = k_temp2*tree_order;
}
}
LAST_STEP:
/* last step, proxies send full data back to the extra ranks */
if( 0 < exchange_node->n_extra_sources &&
root == exchange_node->rank_extra_sources_array[0]) {
/* regardless, I will bump the ready flag and set it in case someone is watching */
/* announce that data is ready */
ready_flag++;
my_ctl_pointer->gflag = ready_flag;
}
FINISHED:
/* debug
fprintf(stderr," my_ctl_pointer->index %d n of this type %d %u \n",
my_ctl_pointer->index,c_input_args->n_of_this_type_in_collective,getpid());
fflush(stderr);
end debug */
my_ctl_pointer->starting_flag_value+=1;
return BCOL_FN_COMPLETE;
}
#endif
#if 0
/* blocking, asynchronous polling gather routine */
int bcol_basesmuma_k_nomial_gather(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
mca_bcol_base_function_t *c_input_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 */
int group_size;
int first_instance=0, flag_offset;
int rc = OMPI_SUCCESS;
int leading_dim, buff_idx, idx;
int *group_list;
int src, comm_src, knt, i, k, j, k_temp1, k_temp2;
int pseudo_root, proxy_root, pseudo_base_adj;
volatile int64_t ready_flag;
int count=input_args->count;
struct ompi_datatype_t* dtype=input_args->dtype;
int root=input_args->root;
int base_adj, base;
int total_peers, my_pow_k;
int probe;
int matched;
int64_t sequence_number=input_args->sequence_num;
mca_bcol_basesmuma_module_t* bcol_module=
(mca_bcol_basesmuma_module_t *)c_input_args->bcol_module;
int my_rank = bcol_module->super.sbgp_partner_module->my_index;
mca_bcol_basesmuma_component_t *cm = &mca_bcol_basesmuma_component;
netpatterns_k_exchange_node_t *exchange_node =
&bcol_module->knomial_allgather_tree;
int buff_offset = bcol_module->super.hier_scather_offset;
/* "indirectors" */
int *list_connected = bcol_module->super.list_n_connected;
int *inv_map = exchange_node->inv_reindex_map;
int *reindex_map = exchange_node->reindex_map;
/*int *reindex_map = exchange_node->reindex_map;*/
/* stray rank == first rank in the extra set */
int stray = exchange_node->k_nomial_stray;
/* tree radix */
int tree_order = exchange_node->tree_order;
/* tree depth */
int pow_k = exchange_node->log_tree_order;
/* largest power of k less than or equal to np */
int cnt = exchange_node->n_largest_pow_tree_order;
/*fprintf(stderr,"tree order %d pow_k %d stray %d root %d\n",tree_order, pow_k, stray, root);*/
/* payload structures */
volatile mca_bcol_basesmuma_payload_t *data_buffs;
volatile char *child_data_pointer;
/* control structures */
volatile mca_bcol_basesmuma_header_t *my_ctl_pointer;
volatile mca_bcol_basesmuma_header_t *child_ctl_pointer;
/*volatile mca_bcol_basesmuma_ctl_struct_t* parent_ctl_pointer; */
size_t pack_len = 0, dt_size;
void *data_addr = (void *)((unsigned char *)input_args->src_desc->data_addr);
/* active in the algorithm */
bool active = true;
#if 0
fprintf(stderr,"Entering sm gather root %d \n",root);
fflush(stderr);
#endif
/* we will work only on packed data - so compute the length*/
/* this is the size of my data, this is not gatherv so it's the same
* for all ranks in the communicator.
*/
ompi_datatype_type_size(dtype, &dt_size);
pack_len=count*dt_size;
/* now set the "real" offset */
buff_offset = buff_offset*pack_len;
buff_idx = input_args->src_desc->buffer_index;
/* Get addressing information */
my_rank = bcol_module->super.sbgp_partner_module->my_index;
/* I have a feeling that I'll need this */
group_list = bcol_module->super.sbgp_partner_module->group_list;
group_size = bcol_module->colls_no_user_data.size_of_group;
leading_dim=bcol_module->colls_no_user_data.size_of_group;
idx=SM_ARRAY_INDEX(leading_dim,buff_idx,0);
/*ctl_structs=(mca_bcol_basesmuma_ctl_struct_t **)
bcol_module->colls_with_user_data.ctl_buffs+idx;
*/
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 = ctl_structs[my_rank]; */
my_ctl_pointer = data_buffs[my_rank].ctl_struct;
/* setup resource recycling */
if( my_ctl_pointer->sequence_number < sequence_number ) {
first_instance=1;
}
if( first_instance ) {
/* Signal arrival */
my_ctl_pointer->flag = -1;
my_ctl_pointer->gflag = -1;
my_ctl_pointer->index=1;
/* this does not need to use any flag values , so only need to
* set the value for subsequent values that may need this */
my_ctl_pointer->starting_flag_value=0;
flag_offset=0;
} else {
/* only one thread at a time will be making progress on this
* collective, so no need to make this atomic */
my_ctl_pointer->index++;
}
/* increment the starting flag by one and return */
flag_offset = my_ctl_pointer->starting_flag_value;
ready_flag = flag_offset + sequence_number + 1;
my_ctl_pointer->sequence_number = sequence_number;
/* debug
fprintf(stderr," sequence_number %lld flag_offset %d starting flag val %d\n",sequence_number,flag_offset, my_ctl_pointer->starting_flag_value);
fflush(stderr);
end debug */
/*
* Fan out from root
*/
/* don't need this either */
/* root is the local leader */
/* calculate the number of steps necessary for this collective */
/* first thing we do is figure out where the root is in our new indexing */
/* find root in new indexing */
pseudo_root = inv_map[root];
/* see if this is larger than the stray */
if( pseudo_root >= stray ) {
/* then we need to define the proxy root, everyone can do this */
proxy_root = pseudo_root - cnt;
}else {
proxy_root = pseudo_root;
}
if( EXTRA_NODE == exchange_node->node_type ) {
/* signal arrival */
my_ctl_pointer->gflag = ready_flag;
/* send is done */
/* poll for data only if I am the root */
/* bump the ready flag */
ready_flag++;
if( root == my_rank ){
/* poll for data from proxy */
src = exchange_node->rank_extra_sources_array[0];
/* get src data buffer */
child_data_pointer = data_buffs[src].payload;
child_ctl_pointer = data_buffs[src].ctl_struct;
/* in this case, you must block */
while(!IS_GDATA_READY(child_ctl_pointer,ready_flag,sequence_number)){
opal_progress();
}
/* receive the data from the proxy, aka pseudo-root */
memcpy((void *) ((unsigned char *) data_addr + buff_offset),
(void *) ((unsigned char *) child_data_pointer+buff_offset)
,pack_len*group_size);
}
goto FINISHED;
} else if( 0 < exchange_node->n_extra_sources ) {
/* am a proxy, poll for pack_len data from extra */
src = exchange_node->rank_extra_sources_array[0];
/* get src data buffer */
child_data_pointer = data_buffs[src].payload;
child_ctl_pointer = data_buffs[src].ctl_struct;
knt = 0;
for( i = 0; i < src; i++){
knt += list_connected[i];
}
/* must block here also */
while(!IS_GDATA_READY(child_ctl_pointer,ready_flag,sequence_number)){
opal_progress();
}
memcpy((void *) ((unsigned char *) data_addr + buff_offset + pack_len*knt),
(void *) ((unsigned char *) child_data_pointer + buff_offset +
pack_len*knt), pack_len*list_connected[src]);
/*fprintf(stderr,"999 proxy received data from %d at offset %d of length %d\n",src,
buff_offset+pack_len*knt,pack_len*list_connected[src]);
*/
}
/* do some figuring */
total_peers = 0;
my_pow_k = pow_k;
k_temp1 = tree_order;
k_temp2 = 1;
for( i = 0; i < pow_k; i++) {
/* then find the base */
/*FIND_BASE(base,my_rank,i+1,tree_order);*/
FIND_BASE(base,exchange_node->reindex_myid,i+1,tree_order);
/* now find the adjusted base */
base_adj = base + (base + proxy_root)%k_temp1;
/* ok, now find out WHO is occupying this slot */
/*pseudo_base_adj = inv_map[base_adj];*/
pseudo_base_adj = reindex_map[base_adj];
if(my_rank == pseudo_base_adj ) {
/* then go ahead and poll for children's data */
for( j = 0; j < (tree_order - 1); j++ ) {
/* send phase
*/
/* get communication partner */
src = exchange_node->rank_exchanges[i][j];
/*fprintf(stderr,"comm_src %d\n",comm_src);*/
/* remember, if we have extra ranks, then we won't participate
* with a least one peer. Make a check
*/
if( src < 0 ){
continue;
}else{
total_peers++;
}
}
} else {
/* announce my arrival */
my_pow_k = i;
break;
}
k_temp1 = k_temp1*tree_order;
k_temp2 = k_temp2*tree_order;
}
if( 0 == my_pow_k ){
/* signal arrival */
my_ctl_pointer->gflag = ready_flag;
goto FINISHED;
}
/* start the k-nomial gather phase */
/* only "active ranks participate, once a rank has forwarded its data, it becomes inactive */
knt = 0;
while(active){
k_temp1 = tree_order;
k_temp2 = 1;
for( i = 0; i < my_pow_k; i++) {
/* then go ahead and poll for children's data */
for( j = 0; j < (tree_order - 1); j++ ) {
matched = 0;
/* send phase
*/
/* get communication partner */
src = exchange_node->rank_exchanges[i][j];
/*fprintf(stderr,"comm_src %d\n",comm_src);*/
/* remember, if we have extra ranks, then we won't participate
* with a least one peer. Make a check
*/
if( src < 0 ){
continue;
}
/*fprintf(stderr,"src %d\n",src);*/
child_data_pointer = data_buffs[src].payload;
child_ctl_pointer = data_buffs[src].ctl_struct;
/* if child has been marked, then skip */
if( sequence_number == child_ctl_pointer->mark ){
continue;
}
for( probe = 0; probe < cm->num_to_probe && (0 == matched); probe++){
if(IS_GDATA_READY(child_ctl_pointer,ready_flag,sequence_number)){
/* mark the child's pointer */
child_ctl_pointer->mark = sequence_number;
/* copy the data */
memcpy((void *) ((unsigned char *) data_addr + buff_offset +
exchange_node->payload_info[i][j].r_offset*pack_len),
(void *) ((unsigned char *) child_data_pointer + buff_offset +
exchange_node->payload_info[i][j].r_offset*pack_len),
exchange_node->payload_info[i][j].r_len*pack_len);
/*
fprintf(stderr,"999 receiving data from %d at offset %d of length %d\n",
exchange_node->rank_exchanges[i][j], buff_offset + exchange_node->payload_info[i][j].r_offset,
exchange_node->payload_info[i][j].r_len*pack_len);
*/
knt++;
if(knt == total_peers) {
/* this is the trick to break the root out,
* only the root should be able to satisfy this
*/
/*
fprintf(stderr,"hello n_actual is %d \n",knt);
fprintf(stderr,"hello n_actual_exch is %d \n",
exchange_node->n_actual_exchanges);
*/
basesmuma: do not call MB directly opal does not always define MB. It is recommended that opal_atomic_[rw]mb is called instead. We will need to address the cases where these functions are no-ops on weak-memory ordered cpus. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30534. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 21:01:57 +04:00
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->gflag = ready_flag;
goto LAST_STEP;
}
matched = 1;
}else{
opal_progress();
}
}
}
}
k_temp1 = k_temp1*tree_order;
k_temp2 = k_temp2*tree_order;
}
LAST_STEP:
/* last step, proxies send full data back to the extra ranks */
if( 0 < exchange_node->n_extra_sources &&
root == exchange_node->rank_extra_sources_array[0]) {
/* regardless, I will bump the ready flag and set it in case someone is watching */
/* announce that data is ready */
ready_flag++;
my_ctl_pointer->gflag = ready_flag;
}
FINISHED:
/* debug
fprintf(stderr," my_ctl_pointer->index %d n of this type %d %u \n",
my_ctl_pointer->index,c_input_args->n_of_this_type_in_collective,getpid());
fflush(stderr);
end debug */
my_ctl_pointer->starting_flag_value+=1;
return BCOL_FN_COMPLETE;
}
#endif
Копировать постоянную ссылку