1
1
openmpi/ompi/mca/coll/ml/coll_ml_inlines.h

638 строки
25 KiB
C
Исходник Обычный вид История

coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2009-2012 Oak Ridge National Laboratory. All rights reserved.
* Copyright (c) 2009-2012 Mellanox Technologies. All rights reserved.
* Copyright (c) 2013 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/** @file */
#ifndef MCA_COLL_ML_INLINES_H
#define MCA_COLL_ML_INLINES_H
#include "ompi_config.h"
BEGIN_C_DECLS
static inline __opal_attribute_always_inline__ int ml_fls(int num)
{
int i = 1;
int j = 0;
if (0 == num) {
return 0;
}
while (i < num) {
i *= 2;
j++;
}
if (i > num) {
j--;
}
return j;
}
static inline __opal_attribute_always_inline__
int mca_coll_ml_buffer_recycling(mca_coll_ml_collective_operation_progress_t *ml_request)
{
mca_coll_ml_module_t *ml_module = (mca_coll_ml_module_t *)ml_request->coll_module;
mca_bcol_base_memory_block_desc_t *ml_memblock = ml_module->payload_block;
uint64_t bank_index = ml_request->fragment_data.buffer_desc->bank_index;
int rc;
opal_atomic_add(&ml_memblock->bank_release_counters[bank_index], 1);
/* Check if the bank is ready for recycling */
if (ml_memblock->bank_release_counters[bank_index] ==
ml_memblock->num_buffers_per_bank ) {
ml_memblock->ready_for_memsync[bank_index] = true;
ML_VERBOSE(10, ("Sync count %d, bank %d", ml_memblock->memsync_counter, bank_index));
assert(ml_memblock->bank_is_busy);
if (ml_memblock->memsync_counter == (int)bank_index) {
while(ml_memblock->ready_for_memsync[ml_memblock->memsync_counter]) {
ML_VERBOSE(10, ("Calling for service barrier: ml_buffer_index - %d %d %d == %d.",
ml_request->fragment_data.buffer_desc->buffer_index,
ml_memblock->memsync_counter,
ml_memblock->bank_release_counters[ml_memblock->memsync_counter],
ml_memblock->num_buffers_per_bank));
/* Setting the ready flag to 0 - unready - done */
ml_memblock->ready_for_memsync[ml_memblock->memsync_counter] = false;
rc = mca_coll_ml_memsync_intra(ml_module, ml_memblock->memsync_counter);
if (OMPI_SUCCESS != rc) {
ML_ERROR(("Failed to start memory sync !!!"));
return rc;
}
opal_atomic_add(&ml_memblock->memsync_counter, 1);
if (ml_memblock->memsync_counter == (int)ml_memblock->num_banks) {
ml_memblock->memsync_counter = 0;
}
ML_VERBOSE(10, ("After service barrier."));
}
} else {
ML_VERBOSE(10, ("Out of order %d", ml_memblock->memsync_counter));
}
}
return OMPI_SUCCESS;
}
static inline __opal_attribute_always_inline__ int coll_ml_fragment_completion_processing(
mca_coll_ml_collective_operation_progress_t *coll_op)
{
/* local variables */
int ret = OMPI_SUCCESS;
size_t bytes_in_this_frag;
struct full_message_t *full_msg_desc = coll_op->fragment_data.message_descriptor;
bool ready_to_release = true, out_of_resource = false;
ML_VERBOSE(10, ("Coll_op %p processing completion", coll_op));
/* Call unpack/pack function */
if (OPAL_LIKELY(NULL != coll_op->process_fn)) {
ret = coll_op->process_fn(coll_op);
switch(ret) {
case OMPI_SUCCESS:
ML_VERBOSE(10, ("unpack done"));
ready_to_release = true;
break;
case ORTE_ERR_NO_MATCH_YET:
ML_VERBOSE(10, ("unexpected packet"));
ready_to_release = false;
break;
default:
ML_ERROR(("Error, unexpected error code %d", ret));
return ret;
}
}
bytes_in_this_frag = coll_op->fragment_data.fragment_size;
ML_VERBOSE(10, ("Delivered %d bytes in frag %d total %d",
full_msg_desc->n_bytes_delivered,
bytes_in_this_frag,
full_msg_desc->n_bytes_total));
/* check for full message completion */
if(full_msg_desc->n_bytes_delivered + bytes_in_this_frag ==
full_msg_desc->n_bytes_total) {
/* message complete - don't update number of bytes delivered, just
* mark the message complete
*/
full_msg_desc->n_bytes_delivered += bytes_in_this_frag;
/* decrement the number of fragments */
full_msg_desc->n_active--;
ML_VERBOSE(10, ("Signaling completion"));
/* here we need to be sure that we point to the first fragment only */
ompi_request_complete(&(coll_op->fragment_data.message_descriptor->super), true);
coll_op->fragment_data.message_descriptor->super.req_status.MPI_ERROR = OMPI_SUCCESS;
} else {
assert(NULL != coll_op->fragment_data.buffer_desc);
/* update the number of bytes delivered */
full_msg_desc->n_bytes_delivered += bytes_in_this_frag;
/* decrement the number of fragments */
full_msg_desc->n_active--;
/* here we need to start the next fragment */
ML_VERBOSE(10, ("Launch frags for %p", coll_op));
if (full_msg_desc->n_bytes_scheduled < full_msg_desc->n_bytes_total) {
ret = coll_op->fragment_data.message_descriptor->fragment_launcher(coll_op);
if (OPAL_UNLIKELY(OMPI_ERR_TEMP_OUT_OF_RESOURCE == ret)) {
out_of_resource = true;
} else if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
ML_VERBOSE(10, ("Failed to launch fragment"));
return ret;
}
}
}
if (ready_to_release) {
/* Check if we have to recycle memory.
* Note: It is safe to recycle ML buffers since the ML buffer data
* already was unpacked to user buffer
*/
if (NULL != coll_op->fragment_data.buffer_desc) {
ret = mca_coll_ml_buffer_recycling(coll_op);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
}
/* if this is not fragment 0, return fragment to the free list.
* fragment 0 will be returned in mca_ml_request_free() which
* is called from the MPI wait() and test() routines.
* We can recover the pointer to the fragement descriptor from
* the MPI level request object, wich is the first element
* in the fragment descriptor.
*/
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
/* I contend that this is a bug. This is not the right way to check
* for the first fragment as it assumes that the first fragment would always
* for any collective have zero as the first offset or that other subsequent
* fragments would not. It is not safe to assume this. The correct check is
* the following one
*/
ML_VERBOSE(10, ("Master ? %p %d", coll_op, coll_op->fragment_data.offset_into_user_buffer));
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
/* This check is in fact a bug. Not the correct definiton of first
* fragment. First fragment is the only fragment that satisfies the
* following criteria
*/
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
/*if (0 != coll_op->fragment_data.offset_into_user_buffer &&
!out_of_resource) {
*/
if (((&coll_op->full_message != coll_op->fragment_data.message_descriptor) &&
!out_of_resource) || IS_COLL_SYNCMEM(coll_op)) {
/* non-zero offset ==> this is not fragment 0 */
CHECK_AND_RECYCLE(coll_op);
}
}
/* return */
return OMPI_SUCCESS;
}
/* task completion */
static inline __opal_attribute_always_inline__ int coll_ml_task_dependency_processing(
mca_coll_ml_task_status_t *task)
{
/* update dependencies */
mca_coll_ml_collective_operation_progress_t *my_schedule_instance =
task->ml_coll_operation;
int n_dependent_tasks = task->rt_num_dependent_tasks;
int dep_task;
for (dep_task = 0; dep_task < n_dependent_tasks; dep_task++)
{
int task_index;
task_index = task->rt_dependent_task_indices[dep_task];
my_schedule_instance->dag_description.status_array[task_index].n_dep_satisfied++;
}
/* return */
return OMPI_SUCCESS;
}
/* collective task completion processing -
* "task" may be removed from list in this routine.
* Thread safety is assumed to be handled outside this routine.
*/
static inline __opal_attribute_always_inline__ int mca_coll_ml_task_completion_processing(
mca_coll_ml_task_status_t **task_status_g, opal_list_t *list)
{
/* local variables */
int ret = OMPI_SUCCESS;
mca_coll_ml_task_status_t *task_status = *task_status_g;
mca_coll_ml_collective_operation_progress_t *coll_op =
task_status->ml_coll_operation;
/* Pasha: Since all our collectives so far use the root
flag, I replacing the call for custom call back function
with setting root_flag.
If we will see that we need some custom functionality,
we will enable it later.
*/
task_status->ml_coll_operation->variable_fn_params.root_flag = true;
#if 0
/* process task completion function,
if any was defined */
if (OPAL_LIKELY(NULL != task_status->task_comp_fn)) {
ret = task_status->task_comp_fn(task_status);
if (ret != OMPI_SUCCESS) {
return ret;
}
}
#endif
/* update dependencies */
ret = coll_ml_task_dependency_processing(task_status);
if (ret != OMPI_SUCCESS) {
ML_VERBOSE(3,("Failed to coll_ml_task_dependency_processing"));
return ret;
}
/* process task completion function,
if any was defined */
if (OPAL_LIKELY(NULL != task_status->task_comp_fn)) {
ret = task_status->task_comp_fn(task_status);
if (ret != OMPI_SUCCESS) {
ML_VERBOSE(3,("Failed to task_comp_fn"));
return ret;
}
}
/* remove the descriptor from the incomplete list
(Pasha: if the list was provided) */
/* No need to put this an any new list - it is associcated
* with the mca_coll_ml_collective_operation_progress_t
* descriptor already
*/
if (NULL != list) {
(*task_status_g) = (mca_coll_ml_task_status_t *)
opal_list_remove_item(list, (opal_list_item_t *)(task_status));
}
/* update completion counter */
coll_op->dag_description.num_tasks_completed++;
if(coll_op->dag_description.num_tasks_completed ==
coll_op->coll_schedule->n_fns)
{
/* the actual fragment descriptor is not on any list, as
* we can get at it from the task descriptors
*/
ret = coll_ml_fragment_completion_processing(coll_op);
if (OMPI_SUCCESS != ret) {
ML_VERBOSE(3,("Failed to coll_ml_fragment_completion_processing"));
return ret;
}
}
/* return */
return ret;
}
static inline __opal_attribute_always_inline__ int mca_coll_ml_generic_collectives_append_to_queue(
mca_coll_ml_collective_operation_progress_t *op_prog,
mca_coll_ml_task_setup_fn_t task_setup)
{
int fn_index;
mca_coll_ml_collective_operation_description_t *op_desc =
op_prog->coll_schedule;
mca_coll_ml_compound_functions_t *func = NULL;
mca_coll_ml_task_status_t *task_status = NULL;
mca_coll_ml_component_t *cm = &mca_coll_ml_component;
ML_VERBOSE(9, ("Calling mca_coll_ml_generic_collectives_launcher"));
/* Init all tasks, before we start them */
for (fn_index = 0; fn_index < op_desc->n_fns; fn_index++) {
func = &op_desc->component_functions[fn_index];
task_status = &op_prog->dag_description.status_array[fn_index];
ML_VERBOSE(9, ("Processing function index %d", fn_index));
assert(NULL != func);
/* Init task status */
task_status->n_dep_satisfied = 0; /* start from zero */
task_status->bcol_fn = func->bcol_function;
/* setup run time parametres */
/* Pasha: do we need the if proctection ? */
if (OPAL_LIKELY(NULL != task_setup)) {
task_setup(task_status, fn_index, func);
}
/* the pointer to operation progress supposed to be set during
construction time. Just want to make sure that it is ok */
assert(task_status->ml_coll_operation == op_prog);
/* We assume that all pointer to functions are defined and it
is not reson to check for null */
assert(NULL != func->bcol_function->coll_fn);
/* In order to preserve ordering on all ranks we have to add it to tail */
/* TBD: Need to review the way we launch fragments */
ML_VERBOSE(9, ("The task %p dependency is %d, appending it on pending list",
(void *)task_status, func->num_dependencies));
OPAL_THREAD_LOCK(&(mca_coll_ml_component.pending_tasks_mutex));
opal_list_append(&cm->pending_tasks, (opal_list_item_t *)task_status);
OPAL_THREAD_UNLOCK(&(mca_coll_ml_component.pending_tasks_mutex));
}
ML_VERBOSE(9, ("Collective was launched !"));
return OMPI_SUCCESS;
}
static inline __opal_attribute_always_inline__ int mca_coll_ml_generic_collectives_launcher(
mca_coll_ml_collective_operation_progress_t *op_prog,
mca_coll_ml_task_setup_fn_t task_setup)
{
int fn_index;
int rc, ret;
mca_coll_ml_collective_operation_description_t *op_desc =
op_prog->coll_schedule;
mca_coll_ml_compound_functions_t *func = NULL;
mca_coll_ml_task_status_t *task_status = NULL;
mca_coll_ml_component_t *cm = &mca_coll_ml_component;
ML_VERBOSE(9, ("Calling mca_coll_ml_generic_collectives_launcher"));
/* Init all tasks, before we start them */
for (fn_index = 0; fn_index < op_desc->n_fns; fn_index++) {
func = &op_desc->component_functions[fn_index];
task_status = &op_prog->dag_description.status_array[fn_index];
ML_VERBOSE(9, ("Processing function index %d", fn_index));
assert(NULL != func);
/* Init task status */
task_status->n_dep_satisfied = 0; /* start from zero */
/* task_status->my_index_in_coll_schedule = fn_index;
pasha: the value is set during init */
task_status->bcol_fn = func->bcol_function;
/* Pasha: disabling support for custom complition functions
task_status->task_comp_fn = func->task_comp_fn;
*/
/* setup run time parametres */
/* Pasha: do we need the if proctection ? */
if (OPAL_LIKELY(NULL != task_setup)) {
task_setup(task_status, fn_index, func);
}
/* the pointer to operation progress supposed to be set during
construction time. Just want to make sure that it is ok */
assert(task_status->ml_coll_operation == op_prog);
/* Task status is done */
/* launch the task and put it on corresponding list (if required) */
/* We assume that all pointer to functions are defined and it
is not reason to check for null */
assert(NULL != func->bcol_function->coll_fn);
}
/* try to start startable */
for (fn_index = 0; fn_index < op_desc->n_fns; fn_index++) {
func = &op_desc->component_functions[fn_index];
task_status = &op_prog->dag_description.status_array[fn_index];
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
/* fire the collective immediately if it has no dependencies */
if (0 == task_status->rt_num_dependencies) {
rc = func->bcol_function->coll_fn(&op_prog->variable_fn_params,
/* Pasha: Need to update the prototype of the func,
right now it is ugly hack for compilation */
(struct mca_bcol_base_function_t *)&func->constant_group_data);
switch(rc) {
case BCOL_FN_NOT_STARTED:
/* put it on pending list */
ML_VERBOSE(9, ("Call to bcol collecitive return BCOL_FN_NOT_STARTED, putting the task on pending list"));
OPAL_THREAD_LOCK(&(mca_coll_ml_component.pending_tasks_mutex));
opal_list_append(&cm->pending_tasks, (opal_list_item_t *)task_status);
OPAL_THREAD_UNLOCK(&(mca_coll_ml_component.pending_tasks_mutex));
break;
case BCOL_FN_STARTED:
/* put it on started list */
ML_VERBOSE(9, ("Call to bcol collecitive return BCOL_FN_STARTED, puting the task on active list"));
OPAL_THREAD_LOCK(&(mca_coll_ml_component.active_tasks_mutex));
opal_list_append(&cm->active_tasks, (opal_list_item_t *)task_status);
OPAL_THREAD_UNLOCK(&(mca_coll_ml_component.active_tasks_mutex));
break;
case BCOL_FN_COMPLETE:
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 task is done ! lets start relevant dependencies */
ML_VERBOSE(9, ("Call to bcol collecitive return BCOL_FN_COMPLETE"));
/* the task does not belong to any list, yes. So passing NULL */
ret = mca_coll_ml_task_completion_processing(&task_status, NULL);
if (OMPI_SUCCESS != ret) {
ML_VERBOSE(9, ("Failed to mca_coll_ml_task_completion_processing"));
return ret;
}
break;
default:
ML_ERROR(("Unknow exit status %d", rc));
return OMPI_ERROR;
}
} else {
/* the task is depend on other, lets put it on pending list */
ML_VERBOSE(9, ("The task %p dependency is %d, putting it on pending list",
(void *)task_status, func->num_dependencies));
OPAL_THREAD_LOCK(&(mca_coll_ml_component.pending_tasks_mutex));
opal_list_append(&cm->pending_tasks, (opal_list_item_t *)task_status);
OPAL_THREAD_UNLOCK(&(mca_coll_ml_component.pending_tasks_mutex));
}
}
ML_VERBOSE(9, ("Collective was launched !"));
return OMPI_SUCCESS;
}
static inline __opal_attribute_always_inline__ mca_coll_ml_collective_operation_progress_t *
mca_coll_ml_alloc_op_prog_single_frag_dag(
mca_coll_ml_module_t *ml_module,
mca_coll_ml_collective_operation_description_t *coll_schedule,
void *src, void *dst, size_t total_bytes,
size_t offset_into_user_buffer
)
{
ompi_free_list_item_t *item;
mca_coll_ml_collective_operation_progress_t *coll_op = NULL;
ompi_request_t *req;
/* Blocking call on fragment allocation (Maybe we want to make it non blocking ?) */
OMPI_FREE_LIST_WAIT_MT(&(ml_module->coll_ml_collective_descriptors),
item);
coll_op = (mca_coll_ml_collective_operation_progress_t *) item;
ML_VERBOSE(10, (">>> Allocating coll op %p", coll_op));
assert(NULL != coll_op);
assert(coll_op->dag_description.status_array[0].item.opal_list_item_refcount == 0);
req = &(coll_op->full_message.super);
OMPI_REQUEST_INIT(req, false);
/* Mark the request ACTIVE. It is critical for MPI_Test()*/
req->req_state = OMPI_REQUEST_ACTIVE;
req->req_status._cancelled = 0;
req->req_status.MPI_ERROR = OMPI_SUCCESS;
MCA_COLL_ML_OP_BASIC_SETUP(coll_op, total_bytes,
offset_into_user_buffer, src, dst, coll_schedule);
/* We do not set sequential, since it is not sequential call */
coll_op->dag_description.num_tasks_completed = 0;
/* Release reference counter have to be zero */
assert(0 == coll_op->pending);
return coll_op;
}
static inline __opal_attribute_always_inline__ mca_coll_ml_collective_operation_progress_t *
mca_coll_ml_duplicate_op_prog_single_frag_dag(
mca_coll_ml_module_t *ml_module,
mca_coll_ml_collective_operation_progress_t *old_op)
{
mca_coll_ml_collective_operation_progress_t *new_op = NULL;
new_op = mca_coll_ml_alloc_op_prog_single_frag_dag(ml_module,
ml_module->coll_ml_bcast_functions[old_op->fragment_data.current_coll_op],
old_op->fragment_data.message_descriptor->dest_user_addr,
old_op->fragment_data.message_descriptor->src_user_addr,
old_op->fragment_data.message_descriptor->n_bytes_total,
old_op->fragment_data.message_descriptor->n_bytes_scheduled);
new_op->fragment_data.current_coll_op = old_op->fragment_data.current_coll_op;
new_op->fragment_data.message_descriptor = old_op->fragment_data.message_descriptor;
return new_op;
}
static inline __opal_attribute_always_inline__ mca_coll_ml_collective_operation_progress_t *
mca_coll_ml_alloc_op_prog_single_frag_seq(
mca_coll_ml_module_t *ml_module,
mca_coll_ml_collective_operation_description_t *coll_schedule,
void *src, void *dst,
size_t total_bytes,
size_t offset_into_user_buffer
)
{
ompi_free_list_item_t *item;
mca_coll_ml_collective_operation_progress_t *coll_op = NULL;
/* Blocking call on fragment allocation (Maybe we want to make it non blocking ?) */
OMPI_FREE_LIST_WAIT_MT(&(ml_module->coll_ml_collective_descriptors),
item);
coll_op = (mca_coll_ml_collective_operation_progress_t *) item;
assert(NULL != coll_op);
MCA_COLL_ML_OP_BASIC_SETUP(coll_op, total_bytes,
offset_into_user_buffer, src, dst, coll_schedule);
/* set sequential data */
/* pasha - do we have something to set ? */
return coll_op;
}
static inline __opal_attribute_always_inline__
void mca_coll_ml_convertor_get_send_frag_size(mca_coll_ml_module_t *ml_module,
size_t *frag_size, struct full_message_t *message_descriptor)
{
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
size_t fragment_size = *frag_size;
opal_convertor_t *dummy_convertor = &message_descriptor->dummy_convertor;
/* The last frag needs special service */
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
if (fragment_size >
(size_t) message_descriptor->send_converter_bytes_packed) {
*frag_size = message_descriptor->send_converter_bytes_packed;
message_descriptor->send_converter_bytes_packed = 0;
return;
}
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
if( (message_descriptor->dummy_conv_position + fragment_size) >
message_descriptor->n_bytes_total ) {
message_descriptor->dummy_conv_position = (message_descriptor->dummy_conv_position + fragment_size)
- message_descriptor->n_bytes_total;
} else {
message_descriptor->dummy_conv_position += fragment_size;
}
opal_convertor_generic_simple_position(dummy_convertor, &message_descriptor->dummy_conv_position);
*frag_size -= dummy_convertor->partial_length;
message_descriptor->send_converter_bytes_packed -= (*frag_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
static inline __opal_attribute_always_inline__ int
mca_coll_ml_launch_sequential_collective (mca_coll_ml_collective_operation_progress_t *coll_op)
{
mca_bcol_base_coll_fn_desc_t *bcol_func;
int ifunc, n_fn, ih, ret;
mca_coll_ml_collective_operation_description_t *sched =
coll_op->coll_schedule;
n_fn = sched->n_fns;
ih = coll_op->sequential_routine.current_active_bcol_fn;
/* if collectives are already pending just add this one to the list */
if (opal_list_get_size (&mca_coll_ml_component.sequential_collectives)) {
opal_list_append(&mca_coll_ml_component.sequential_collectives, (opal_list_item_t *) coll_op);
return OMPI_SUCCESS;
}
for (ifunc = ih; ifunc < n_fn; ifunc++, coll_op->sequential_routine.current_active_bcol_fn++) {
ret = coll_op->sequential_routine.seq_task_setup(coll_op);
if (OMPI_SUCCESS != ret) {
return ret;
}
bcol_func = (sched->component_functions[ifunc].bcol_function);
ret = bcol_func->coll_fn(&coll_op->variable_fn_params,
(struct mca_bcol_base_function_t *) &sched->component_functions[ifunc].constant_group_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
if (BCOL_FN_COMPLETE == ret) {
if (ifunc == n_fn - 1) {
ret = coll_ml_fragment_completion_processing(coll_op);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
mca_coll_ml_abort_ml("Failed to run coll_ml_fragment_completion_processing");
}
return OMPI_SUCCESS;
}
} else {
if (BCOL_FN_STARTED == ret) {
coll_op->sequential_routine.current_bcol_status = SEQ_TASK_IN_PROG;
} else {
coll_op->sequential_routine.current_bcol_status = SEQ_TASK_PENDING;
}
ML_VERBOSE(10, ("Adding pending bcol to the progress list to access by ml_progress func-id %d", ifunc));
opal_list_append(&mca_coll_ml_component.sequential_collectives, (opal_list_item_t *) coll_op);
break;
}
}
return OMPI_SUCCESS;
}
END_C_DECLS
#endif