ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
51c5daf1b4
openmpi/ompi/mca/bcol/basesmuma/bcol_basesmuma_setup.c

591 строка
20 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 -*- */
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
/*
* Copyright (c) 2009-2012 Oak Ridge National Laboratory. All rights reserved.
* Copyright (c) 2009-2012 Mellanox Technologies. All rights reserved.
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
* Copyright (c) 2013-2014 Los Alamos National Security, LLC.
Move the RTE framework change into the trunk. With this change, all non-CR runtime code goes through one of the rte, dpm, or pubsub frameworks. This commit was SVN r27934.
2013-01-28 03:25:10 +04:00
* All rights reserved.
s/CACHE_LINE_SIZE/BASESMUMA_CACHE_LINE_SIZE/g to avoid a system macro name clash on some BSDs. cmr=v1.7.4:reviewer=pasha This commit was SVN r30230.
2014-01-10 20:48:43 +04:00
* Copyright (c) 2014 Cisco Systems, Inc. All rights reserved.
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*
*/
#include "ompi_config.h"
#include "mpi.h"
#include "ompi/constants.h"
#include "ompi/communicator/communicator.h"
#include "ompi/mca/mpool/base/base.h"
#include "ompi/mca/bcol/bcol.h"
#include "ompi/mca/bcol/base/base.h"
Moving mca/common/netpatterns and commpaterns to ompi/patterns. This commit was SVN r28035.
2013-02-06 01:52:55 +04:00
#include "ompi/patterns/comm/coll_ops.h"
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
#include "opal/class/opal_object.h"
#include "opal/dss/dss.h"
#include "bcol_basesmuma.h"
int base_bcol_basesmuma_setup_ctl_struct(
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
mca_bcol_basesmuma_module_t *sm_bcol_module,
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
mca_bcol_basesmuma_component_t *cs,
sm_buffer_mgmt *ctl_mgmt);
/* this is the new one, uses the pml allgather */
int base_bcol_basesmuma_exchange_offsets(
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
mca_bcol_basesmuma_module_t *sm_bcol_module,
void **result_array, uint64_t mem_offset, int loop_limit,
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
int leading_dim)
{
int ret=OMPI_SUCCESS,i;
int count;
int index_in_group;
char *send_buff;
char *recv_buff;
uint64_t rem_mem_offset;
/* malloc some memory */
count = sizeof(uint64_t) + sizeof(int);
send_buff = (char *) malloc(count);
recv_buff = (char *) malloc(count *
sm_bcol_module->super.sbgp_partner_module->group_size);
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
/* exchange the base pointer for the controls structures - gather
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
* every one else's infromation.
*/
/* pack the offset of the allocated region */
memcpy((void *) send_buff, (void *) &(sm_bcol_module->super.sbgp_partner_module->my_index), sizeof(int));
memcpy((void *) (send_buff+ sizeof(int)), (void *) &(mem_offset), sizeof(uint64_t));
/* get the offsets from all procs, so can setup the control data
* structures.
*/
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
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
ret=comm_allgather_pml((void *) send_buff,(void *) recv_buff,count,
MPI_BYTE,
sm_bcol_module->super.sbgp_partner_module->my_index,
sm_bcol_module->super.sbgp_partner_module->group_size,
sm_bcol_module->super.sbgp_partner_module->group_list,
sm_bcol_module->super.sbgp_partner_module->group_comm);
if( OMPI_SUCCESS != ret ) {
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
/* get the control stucture offsets within the shared memory
* region and populate the control structures - we do not assume
* any symmetry in memory layout of each process
*/
/* loop over the procs in the group */
for(i = 0; i < sm_bcol_module->super.sbgp_partner_module->group_size; i++){
int array_id;
/* get this peer's index in the group */
memcpy((void *) &index_in_group, (void *) (recv_buff + i*count) , sizeof(int));
/* get the offset */
memcpy((void *) &rem_mem_offset, (void *) (recv_buff + i*count + sizeof(int)), sizeof(uint64_t));
array_id=SM_ARRAY_INDEX(leading_dim,0,index_in_group);
Cleanup warnings found when building v1.7 cmr:v1.7 This commit was SVN r28279.
2013-04-03 21:37:02 +04:00
result_array[array_id]=(void *)(uintptr_t)rem_mem_offset;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
exit_ERROR:
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
/* clean up */
if( NULL != send_buff ) {
free(send_buff);
send_buff = NULL;
}
if( NULL != recv_buff ) {
free(recv_buff);
recv_buff = NULL;
}
return ret;
}
#if 0
int base_bcol_basesmuma_exchange_offsets(
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
mca_bcol_basesmuma_module_t *sm_bcol_module,
void **result_array, uint64_t mem_offset, int loop_limit,
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
int leading_dim)
{
int ret=OMPI_SUCCESS,i,dummy;
int index_in_group, pcnt;
opal_list_t peers;
Move the RTE framework change into the trunk. With this change, all non-CR runtime code goes through one of the rte, dpm, or pubsub frameworks. This commit was SVN r27934.
2013-01-28 03:25:10 +04:00
ompi_namelist_t *peer;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
ompi_proc_t *proc_temp, *my_id;
opal_buffer_t *send_buffer = OBJ_NEW(opal_buffer_t);
opal_buffer_t *recv_buffer = OBJ_NEW(opal_buffer_t);
uint64_t rem_mem_offset;
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
/* exchange the base pointer for the controls structures - gather
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
* every one else's infromation.
*/
/* get list of procs that will participate in the communication */
OBJ_CONSTRUCT(&peers, opal_list_t);
for (i = 0; i < sm_bcol_module->super.sbgp_partner_module->group_size; i++) {
/* get the proc info */
proc_temp = ompi_comm_peer_lookup(
sm_bcol_module->super.sbgp_partner_module->group_comm,
sm_bcol_module->super.sbgp_partner_module->group_list[i]);
Move the RTE framework change into the trunk. With this change, all non-CR runtime code goes through one of the rte, dpm, or pubsub frameworks. This commit was SVN r27934.
2013-01-28 03:25:10 +04:00
peer = OBJ_NEW(ompi_namelist_t);
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
peer->name.jobid = proc_temp->proc_name.jobid;
peer->name.vpid = proc_temp->proc_name.vpid;
Move the RTE framework change into the trunk. With this change, all non-CR runtime code goes through one of the rte, dpm, or pubsub frameworks. This commit was SVN r27934.
2013-01-28 03:25:10 +04:00
opal_list_append(&peers,&peer->super); /* this is with the new field called "super" in ompi_namelist_t struct */
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
/* pack up the data into the allgather send buffer */
if (NULL == send_buffer || NULL == recv_buffer) {
bcol/basesmuma: use framework output for error message and fix a rounding error. cmr=v1.7.5:reviewer=pasha This commit was SVN r30929.
2014-03-04 20:55:57 +04:00
opal_output (ompi_bcol_base_framework.framework_output, "Cannot allocate memory for sbuffer or rbuffer\n");
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
ret = OMPI_ERROR;
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
/* get my proc information */
my_id = ompi_proc_local();
/* pack my information */
ret = opal_dss.pack(send_buffer,
&(sm_bcol_module->super.sbgp_partner_module->my_index),1,OPAL_UINT32);
Move the RTE framework change into the trunk. With this change, all non-CR runtime code goes through one of the rte, dpm, or pubsub frameworks. This commit was SVN r27934.
2013-01-28 03:25:10 +04:00
if (OMPI_SUCCESS != ret) {
bcol/basesmuma: use framework output for error message and fix a rounding error. cmr=v1.7.5:reviewer=pasha This commit was SVN r30929.
2014-03-04 20:55:57 +04:00
opal_output (ompi_bcol_base_framework.framework_output, "Error packing my_index!!\n");
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
/* pack the offset of the allocated region */
ret = opal_dss.pack(send_buffer,&(mem_offset),1,OPAL_UINT64);
Move the RTE framework change into the trunk. With this change, all non-CR runtime code goes through one of the rte, dpm, or pubsub frameworks. This commit was SVN r27934.
2013-01-28 03:25:10 +04:00
if (OMPI_SUCCESS != ret) {
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
/* get the offsets from all procs, so can setup the control data
* structures.
*/
Move the RTE framework change into the trunk. With this change, all non-CR runtime code goes through one of the rte, dpm, or pubsub frameworks. This commit was SVN r27934.
2013-01-28 03:25:10 +04:00
if (OMPI_SUCCESS != (ret = ompi_rte_allgather_list(&peers, send_buffer, recv_buffer))) {
bcol/basesmuma: use framework output for error message and fix a rounding error. cmr=v1.7.5:reviewer=pasha This commit was SVN r30929.
2014-03-04 20:55:57 +04:00
opal_output (ompi_bcol_base_framework.framework_output, "ompi_rte_allgather_list returned error %d\n", ret);
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
/* unpack the dummy */
pcnt=1;
ret = opal_dss.unpack(recv_buffer,&dummy, &pcnt, OPAL_INT32);
Move the RTE framework change into the trunk. With this change, all non-CR runtime code goes through one of the rte, dpm, or pubsub frameworks. This commit was SVN r27934.
2013-01-28 03:25:10 +04:00
if (OMPI_SUCCESS != ret) {
bcol/basesmuma: use framework output for error message and fix a rounding error. cmr=v1.7.5:reviewer=pasha This commit was SVN r30929.
2014-03-04 20:55:57 +04:00
opal_output (ompi_bcol_base_framework.framework_output, "unpack returned error %d for dummy\n",ret);
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
/* get the control stucture offsets within the shared memory
* region and populate the control structures - we do not assume
* any symmetry in memory layout of each process
*/
/* loop over the procs in the group */
for(i = 0; i < sm_bcol_module->super.sbgp_partner_module->group_size; i++){
int array_id;
pcnt=1;
ret = opal_dss.unpack(recv_buffer,&index_in_group, &pcnt, OPAL_UINT32);
Move the RTE framework change into the trunk. With this change, all non-CR runtime code goes through one of the rte, dpm, or pubsub frameworks. This commit was SVN r27934.
2013-01-28 03:25:10 +04:00
if (OMPI_SUCCESS != ret) {
bcol/basesmuma: use framework output for error message and fix a rounding error. cmr=v1.7.5:reviewer=pasha This commit was SVN r30929.
2014-03-04 20:55:57 +04:00
opal_output (ompi_bcol_base_framework.framework_output, "unpack returned error %d for remote index_in_group\n",ret);
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
/* get the offset */
pcnt=1;
ret = opal_dss.unpack(recv_buffer,&rem_mem_offset, &pcnt, OPAL_UINT64);
Move the RTE framework change into the trunk. With this change, all non-CR runtime code goes through one of the rte, dpm, or pubsub frameworks. This commit was SVN r27934.
2013-01-28 03:25:10 +04:00
if (OMPI_SUCCESS != ret) {
bcol/basesmuma: use framework output for error message and fix a rounding error. cmr=v1.7.5:reviewer=pasha This commit was SVN r30929.
2014-03-04 20:55:57 +04:00
opal_output (ompi_bcol_base_framework.framework_output, "unpack returned error %d for remote memory offset\n",ret);
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
array_id=SM_ARRAY_INDEX(leading_dim,0,index_in_group);
result_array[array_id]=(void *)rem_mem_offset;
}
/* clean up */
Move the RTE framework change into the trunk. With this change, all non-CR runtime code goes through one of the rte, dpm, or pubsub frameworks. This commit was SVN r27934.
2013-01-28 03:25:10 +04:00
peer=(ompi_namelist_t *)opal_list_remove_first(&peers);
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
while( NULL !=peer) {
OBJ_RELEASE(peer);
Move the RTE framework change into the trunk. With this change, all non-CR runtime code goes through one of the rte, dpm, or pubsub frameworks. This commit was SVN r27934.
2013-01-28 03:25:10 +04:00
peer=(ompi_namelist_t *)opal_list_remove_first(&peers);
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
OBJ_DESTRUCT(&peers);
if( send_buffer ) {
OBJ_RELEASE(send_buffer);
}
if( recv_buffer ) {
OBJ_RELEASE(recv_buffer);
}
return ret;
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
exit_ERROR:
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
/* free peer list */
Move the RTE framework change into the trunk. With this change, all non-CR runtime code goes through one of the rte, dpm, or pubsub frameworks. This commit was SVN r27934.
2013-01-28 03:25:10 +04:00
peer=(ompi_namelist_t *)opal_list_remove_first(&peers);
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
while( NULL !=peer) {
OBJ_RELEASE(peer);
Move the RTE framework change into the trunk. With this change, all non-CR runtime code goes through one of the rte, dpm, or pubsub frameworks. This commit was SVN r27934.
2013-01-28 03:25:10 +04:00
peer=(ompi_namelist_t *)opal_list_remove_first(&peers);
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
OBJ_DESTRUCT(&peers);
if( send_buffer ) {
OBJ_RELEASE(send_buffer);
}
if( recv_buffer ) {
OBJ_RELEASE(recv_buffer);
}
return ret;
}
#endif
static int base_bcol_basesmuma_exchange_ctl_params(
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
mca_bcol_basesmuma_module_t *sm_bcol_module,
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
mca_bcol_basesmuma_component_t *cs,
sm_buffer_mgmt *ctl_mgmt, list_data_t *data_blk)
{
int ret=OMPI_SUCCESS,i,loop_limit;
int leading_dim, buf_id;
bcol/basesmuma: fix initialization on 32-bit platforms The initialization code did several allgathers on void *'s using MPI_LONG_LONG_INT. This will produce the wrong result on 32-bit platforms. Instead use MPI_BYTE with count = sizeof (void *). cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30627. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-08 04:00:30 +04:00
void *mem_offset;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
unsigned char *base_ptr;
mca_bcol_basesmuma_ctl_struct_t *ctl_ptr;
/* data block base offset in the mapped file */
bcol/basesmuma: fix initialization on 32-bit platforms The initialization code did several allgathers on void *'s using MPI_LONG_LONG_INT. This will produce the wrong result on 32-bit platforms. Instead use MPI_BYTE with count = sizeof (void *). cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30627. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-08 04:00:30 +04:00
mem_offset = (void *)((uintptr_t)data_blk->data -
(uintptr_t)cs->sm_ctl_structs->data_addr);
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
/* number of buffers in data block */
loop_limit=cs->basesmuma_num_mem_banks+ctl_mgmt->number_of_buffs;
leading_dim=ctl_mgmt->size_of_group;
bcol/basesmuma: fix initialization on 32-bit platforms The initialization code did several allgathers on void *'s using MPI_LONG_LONG_INT. This will produce the wrong result on 32-bit platforms. Instead use MPI_BYTE with count = sizeof (void *). cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30627. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-08 04:00:30 +04:00
ret=comm_allgather_pml(&mem_offset, ctl_mgmt->ctl_buffs, sizeof(void *),
MPI_BYTE, sm_bcol_module->super.sbgp_partner_module->my_index,
sm_bcol_module->super.sbgp_partner_module->group_size,
sm_bcol_module->super.sbgp_partner_module->group_list,
sm_bcol_module->super.sbgp_partner_module->group_comm);
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
if( OMPI_SUCCESS != ret ) {
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
#if 0
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
ret=base_bcol_basesmuma_exchange_offsets( sm_bcol_module,
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
(void **)ctl_mgmt->ctl_buffs, mem_offset, loop_limit, leading_dim);
if( OMPI_SUCCESS != ret ) {
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
#endif
/* convert memory offset to virtual address in current rank */
for (i=0;i< sm_bcol_module->super.sbgp_partner_module->group_size;i++) {
/* get the base pointer */
int array_id=SM_ARRAY_INDEX(leading_dim,0,i);
if( i == sm_bcol_module->super.sbgp_partner_module->my_index) {
/* me */
base_ptr=cs->sm_ctl_structs->map_addr;
} else {
base_ptr=sm_bcol_module->ctl_backing_files_info[i]->sm_mmap->map_addr;
}
ctl_mgmt->ctl_buffs[array_id]=(void *)
Cleanup warnings found when building v1.7 cmr:v1.7 This commit was SVN r28279.
2013-04-03 21:37:02 +04:00
(uintptr_t)(((uint64_t)(uintptr_t)ctl_mgmt->ctl_buffs[array_id])+(uint64_t)(uintptr_t)base_ptr);
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
for( buf_id = 1 ; buf_id < loop_limit ; buf_id++ ) {
int array_id_m1=SM_ARRAY_INDEX(leading_dim,(buf_id-1),i);
array_id=SM_ARRAY_INDEX(leading_dim,buf_id,i);
Cleanup warnings found when building v1.7 cmr:v1.7 This commit was SVN r28279.
2013-04-03 21:37:02 +04:00
ctl_mgmt->ctl_buffs[array_id]=(void *) (uintptr_t)((uint64_t)(uintptr_t)(ctl_mgmt->ctl_buffs[array_id_m1])+
(uint64_t)(uintptr_t)sizeof(mca_bcol_basesmuma_ctl_struct_t));
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
}
/* initialize my control structues */
for( buf_id = 0 ; buf_id < loop_limit ; buf_id++ ) {
int my_idx=sm_bcol_module->super.sbgp_partner_module->my_index;
int array_id=SM_ARRAY_INDEX(leading_dim,buf_id,my_idx);
ctl_ptr = (mca_bcol_basesmuma_ctl_struct_t *)
ctl_mgmt->ctl_buffs[array_id];
/* initialize the data structures - RLG, this is only one data
* structure that needs to be initialized, more are missing */
ctl_ptr->sequence_number=-1;
ctl_ptr->flag=-1;
ctl_ptr->index=0;
ctl_ptr->src_ptr = NULL;
}
return ret;
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
exit_ERROR:
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
return ret;
}
int base_bcol_basesmuma_setup_ctl_struct(
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
mca_bcol_basesmuma_module_t *sm_bcol_module,
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
mca_bcol_basesmuma_component_t *cs,
sm_buffer_mgmt *ctl_mgmt)
{
int ret=OMPI_SUCCESS,i,n_ctl,n_levels;
int n_ctl_structs;
int cnt;
size_t malloc_size;
bcol_basesmuma_smcm_file_t input_file;
/*
* set my no user-data conrol structures
*/
/* number of banks and regions per bank are already a power of 2 */
n_ctl_structs=cs->basesmuma_num_mem_banks*
cs->basesmuma_num_regions_per_bank;
/* initialize the control structure management struct -
* for collectives without user data
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
*---------------------------------------------------------------
*/
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
ctl_mgmt->number_of_buffs=n_ctl_structs;
ctl_mgmt->num_mem_banks=
cs->basesmuma_num_mem_banks;
ctl_mgmt->num_buffs_per_mem_bank=
cs->basesmuma_num_regions_per_bank;
ctl_mgmt->size_of_group=
sm_bcol_module->super.sbgp_partner_module->group_size;
roundup_to_power_radix(2,cs->basesmuma_num_regions_per_bank,&n_levels);
ctl_mgmt->log2_num_buffs_per_mem_bank=n_levels;
roundup_to_power_radix(2,n_ctl_structs,&n_levels);
ctl_mgmt->log2_number_of_buffs=n_levels;
ctl_mgmt->mask=n_ctl_structs-1;
sm_bcol_module->super.n_poll_loops=cs->n_poll_loops;
malloc_size=
(ctl_mgmt->number_of_buffs +
ctl_mgmt->num_mem_banks ) *
ctl_mgmt->size_of_group *
sizeof(void *);
ctl_mgmt->ctl_buffs= malloc(malloc_size);
if( !ctl_mgmt->ctl_buffs ) {
ret=OMPI_ERR_OUT_OF_RESOURCE;
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
/* exchange remote addressing information */
input_file.file_name=cs->sm_ctl_structs->map_path;
input_file.size=cs->sm_ctl_structs->map_size;
input_file.size_ctl_structure=0;
s/CACHE_LINE_SIZE/BASESMUMA_CACHE_LINE_SIZE/g to avoid a system macro name clash on some BSDs. cmr=v1.7.4:reviewer=pasha This commit was SVN r30230.
2014-01-10 20:48:43 +04:00
input_file.data_seg_alignment=BASESMUMA_CACHE_LINE_SIZE;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
input_file.mpool_size=cs->sm_ctl_structs->map_size;
ret=bcol_basesmuma_smcm_allgather_connection(
sm_bcol_module,
sm_bcol_module->super.sbgp_partner_module,
&(cs->sm_connections_list),
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
&(sm_bcol_module->ctl_backing_files_info),
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
sm_bcol_module->super.sbgp_partner_module->group_comm,
input_file, cs->clt_base_fname,
false);
bcol/basesmuma: check the return code of bcol_basesmuma_smcm_allgather_connection. Fixes a segmentation fault found by the bogus intercomm_create test. cmr=v1.7.4:review=manjugv This commit was SVN r30527.
2014-02-01 02:20:25 +04:00
if (OMPI_SUCCESS != ret) {
goto exit_ERROR;
}
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
/* fill in the pointer to other ranks scartch shared memory */
sm_bcol_module->shared_memory_scratch_space=
malloc(sizeof(void *)*
sm_bcol_module->super.sbgp_partner_module->group_size);
if( !sm_bcol_module->shared_memory_scratch_space ) {
bcol/basesmuma: use framework output for error message and fix a rounding error. cmr=v1.7.5:reviewer=pasha This commit was SVN r30929.
2014-03-04 20:55:57 +04:00
opal_output (ompi_bcol_base_framework.framework_output, "Cannot allocate memory for shared_memory_scratch_space.\n");
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
ret = OMPI_ERR_OUT_OF_RESOURCE;
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
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
for(i=0 ; i < sm_bcol_module->super.sbgp_partner_module->group_size ; i++ ) {
if(i == sm_bcol_module->super.sbgp_partner_module->my_index) {
/* local file data is not cached in thi slist */
continue;
}
sm_bcol_module->shared_memory_scratch_space[i]=(void *)(
(char *)(sm_bcol_module->ctl_backing_files_info[i]->sm_mmap)+
cs->scratch_offset_from_base_ctl_file);
}
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
i=sm_bcol_module->super.sbgp_partner_module->my_index;
sm_bcol_module->shared_memory_scratch_space[i]=(void *)(
(char *)(cs->sm_ctl_structs->map_addr)+cs->scratch_offset_from_base_ctl_file);
/*
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
* setup the no-data buffer managment data
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
*/
n_ctl=ctl_mgmt->num_mem_banks;
ctl_mgmt->ctl_buffs_mgmt=(mem_bank_management_t *)
malloc(sizeof(mem_bank_management_t)*n_ctl);
if( !ctl_mgmt->ctl_buffs_mgmt ) {
bcol/basesmuma: use framework output for error message and fix a rounding error. cmr=v1.7.5:reviewer=pasha This commit was SVN r30929.
2014-03-04 20:55:57 +04:00
opal_output (ompi_bcol_base_framework.framework_output, "Cannot allocate memory for ctl_buffs_mgmt. ret = %d\n",ret);
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
ret = OMPI_ERROR;
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
/* initialize each individual element */
cnt=cs->basesmuma_num_regions_per_bank*cs->basesmuma_num_mem_banks;
for( i=0 ; i < n_ctl ; i++ ) {
opal_list_item_t *item;
opal_mutex_t *mutex_ptr;
ctl_mgmt->ctl_buffs_mgmt[i].bank_gen_counter= 0;
ctl_mgmt->ctl_buffs_mgmt[i].available_buffers=
ctl_mgmt->num_buffs_per_mem_bank;
ctl_mgmt->ctl_buffs_mgmt[i].number_of_buffers=
ctl_mgmt->num_buffs_per_mem_bank;
ctl_mgmt->ctl_buffs_mgmt[i].n_buffs_freed= 0;
mutex_ptr= &(ctl_mgmt->ctl_buffs_mgmt[i].mutex);
OBJ_CONSTRUCT(mutex_ptr, opal_mutex_t);
ctl_mgmt->ctl_buffs_mgmt[i].index_shared_mem_ctl_structs=i;
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
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
item=(opal_list_item_t *)&(ctl_mgmt->ctl_buffs_mgmt[i].nb_barrier_desc);
OBJ_CONSTRUCT(item,opal_list_item_t);
ctl_mgmt->ctl_buffs_mgmt[i].nb_barrier_desc.sm_module=
sm_bcol_module;
ctl_mgmt->ctl_buffs_mgmt[i].nb_barrier_desc.pool_index= i;
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
/* get the sm_buffer_mgmt pointer for the control structures */
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
ctl_mgmt->ctl_buffs_mgmt[i].nb_barrier_desc.coll_buff=ctl_mgmt;
ctl_mgmt->ctl_buffs_mgmt[i].nb_barrier_desc.ml_memory_block_descriptor=
NULL;
cnt++;
}
return ret;
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
exit_ERROR:
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
return ret;
}
/*
* this function initializes the internal scratch buffers and control
* structures that will be used by the module. It also intitializes
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
* the payload buffer management structures.
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
*/
int base_bcol_basesmuma_setup_library_buffers(
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
mca_bcol_basesmuma_module_t *sm_bcol_module,
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
mca_bcol_basesmuma_component_t *cs)
{
int ret=OMPI_SUCCESS,i;
int n_ctl_structs;
size_t ctl_segement_size,total_memory;
int max_elements;
unsigned char *data_ptr;
/* */
/* setup the control struct memory */
if(!cs->sm_ctl_structs) {
ret = mca_bcol_basesmuma_allocate_sm_ctl_memory(cs);
if(OMPI_SUCCESS != ret) {
bcol/basesmuma: use framework output for error message and fix a rounding error. cmr=v1.7.5:reviewer=pasha This commit was SVN r30929.
2014-03-04 20:55:57 +04:00
opal_output (ompi_bcol_base_framework.framework_output, "In bcol_comm_query mca_bcol_basesmuma_allocate_sm_ctl_memory failed\n");
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
return ret;
}
/*
* put the memory onto the free list - we have worried about
* alignment in the mpool allocation, and assume that the
* ctl structures have the approriate size to mantain alignment
*/
/* figure out segment size */
n_ctl_structs=cs->basesmuma_num_mem_banks*
cs->basesmuma_num_regions_per_bank;
/* add memory for the control structure used for recycling the banks */
n_ctl_structs+=cs->basesmuma_num_mem_banks;
ctl_segement_size=n_ctl_structs*
sizeof(mca_bcol_basesmuma_ctl_struct_t);
total_memory=cs->sm_ctl_structs->map_size - (
(char *)(cs->sm_ctl_structs->data_addr)-
(char *)(cs->sm_ctl_structs->map_addr));
total_memory-=cs->my_scratch_shared_memory_size;
max_elements=total_memory/ctl_segement_size;
/* populate the free list */
data_ptr=cs->sm_ctl_structs->data_addr;
for( i=0 ; i < max_elements ; i++ ) {
list_data_t *item=OBJ_NEW(list_data_t);
if( !item ) {
ret=OMPI_ERR_OUT_OF_RESOURCE;
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
item->data=(void *)data_ptr;
opal_list_append(&(cs->ctl_structures),(opal_list_item_t *)item);
data_ptr+=ctl_segement_size;
}
/* set the scratch memory pointer and offset */
cs->my_scratch_shared_memory=(char *)data_ptr;
cs->scratch_offset_from_base_ctl_file=(size_t)
((char *)data_ptr-(char *)cs->sm_ctl_structs->map_addr);
/* At this stage the memory is mapped and ready to use by the local rank.
* However, the memory of other processes has not yet been mmaped into the
* memory of this process.
*/
}
/* intialize no_userdata_ctl */
sm_bcol_module->no_userdata_ctl=(list_data_t *)
opal_list_remove_last(&(cs->ctl_structures));
if( !sm_bcol_module->no_userdata_ctl) {
ret=OMPI_ERR_OUT_OF_RESOURCE;
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
ret=base_bcol_basesmuma_setup_ctl_struct(
sm_bcol_module, cs, &(sm_bcol_module->colls_no_user_data));
if( OMPI_SUCCESS != ret ) {
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
/* intialize userdata_ctl */
sm_bcol_module->userdata_ctl=(list_data_t *)
opal_list_remove_last(&(cs->ctl_structures));
if( !sm_bcol_module->userdata_ctl) {
ret=OMPI_ERR_OUT_OF_RESOURCE;
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
ret=base_bcol_basesmuma_setup_ctl_struct(
sm_bcol_module, cs, &(sm_bcol_module->colls_with_user_data));
if( OMPI_SUCCESS != ret ) {
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
/* used for blocking recursive doubling barrier */
sm_bcol_module->index_blocking_barrier_memory_bank=0;
/* gather the offsets of the control structs relative to the base
* of the shared memory file, and fill in the table with the
* address of all the control structues.
*/
ret= base_bcol_basesmuma_exchange_ctl_params(sm_bcol_module, cs,
&(sm_bcol_module->colls_no_user_data),sm_bcol_module->no_userdata_ctl);
if( OMPI_SUCCESS != ret ) {
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
ret= base_bcol_basesmuma_exchange_ctl_params(sm_bcol_module, cs,
&(sm_bcol_module->colls_with_user_data),sm_bcol_module->userdata_ctl);
if( OMPI_SUCCESS != ret ) {
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
goto exit_ERROR;
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
}
return ret;
Per Marco, rename ERROR tags to exit_ERROR to avoid cygwin reserved name issues. Refs trac:4085 This commit was SVN r30239. The following Trac tickets were found above: Ticket 4085 --> https://svn.open-mpi.org/trac/ompi/ticket/4085
2014-01-10 22:00:49 +04:00
exit_ERROR:
Adding Hierarchical Collectives project to the Open MPI trunk. The project includes following components and frameworks: - ML Collective component - NETPATTERNS and COMMPATTERNS common components - BCOL framework - SBGP framework Note: By default the ML collective component is disabled. In order to enable new collectives user should bump up the priority of ml component (coll_ml_priority) ============================================= Primary Contributors (in alphabetical order): Ishai Rabinovich (Mellanox) Joshua S. Ladd (ORNL / Mellanox) Manjunath Gorentla Venkata (ORNL) Mike Dubman (Mellanox) Noam Bloch (Mellanox) Pavel (Pasha) Shamis (ORNL / Mellanox) Richard Graham (ORNL / Mellanox) Vasily Filipov (Mellanox) This commit was SVN r27078.
2012-08-16 23:11:35 +04:00
return ret;
}
OBJ_CLASS_INSTANCE(list_data_t,
opal_list_item_t, NULL, NULL);
Копировать постоянную ссылку