ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
c3df744a3b
openmpi/ompi/mca/coll/ml/coll_ml_allreduce.c

552 строки
22 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 Los Alamos National Security, LLC. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/** @file */
#include "ompi_config.h"
#include <stdlib.h>
#include "ompi/constants.h"
#include "opal/threads/mutex.h"
#include "ompi/communicator/communicator.h"
#include "ompi/mca/coll/coll.h"
#include "ompi/mca/bcol/bcol.h"
#include "opal/sys/atomic.h"
#include "coll_ml.h"
#include "coll_ml_select.h"
#include "coll_ml_allocation.h"
static int mca_coll_ml_allreduce_small_unpack(mca_coll_ml_collective_operation_progress_t *coll_op)
{
int ret;
/* need to put in more */
int count = coll_op->variable_fn_params.count;
ompi_datatype_t *dtype = coll_op->variable_fn_params.dtype;
void *dest = (void *)((uintptr_t)coll_op->full_message.dest_user_addr +
(uintptr_t)coll_op->fragment_data.offset_into_user_buffer);
void *src = (void *)((uintptr_t)coll_op->fragment_data.buffer_desc->data_addr +
(size_t)coll_op->variable_fn_params.rbuf_offset);
ret = ompi_datatype_copy_content_same_ddt(dtype, (int32_t) count, (char *) dest,
(char *) src);
if (ret < 0) {
return OMPI_ERROR;
}
ML_VERBOSE(10, ("sbuf addr %p, sbuf offset %d, rbuf addr %p, rbuf offset %d.",
src, coll_op->variable_fn_params.sbuf_offset, dest,
coll_op->variable_fn_params.rbuf_offset));
return OMPI_SUCCESS;
}
static int mca_coll_ml_allreduce_task_setup(mca_coll_ml_collective_operation_progress_t *coll_op)
{
int fn_idx, h_level, my_index, root;
mca_sbgp_base_module_t *sbgp;
mca_coll_ml_topology_t *topo = coll_op->coll_schedule->topo_info;
fn_idx = coll_op->sequential_routine.current_active_bcol_fn;
h_level = coll_op->coll_schedule->component_functions[fn_idx].h_level;
sbgp = topo->component_pairs[h_level].subgroup_module;
my_index = sbgp->my_index;
/* In the case of allreduce, the local leader is always the root */
root = 0;
if (my_index == root) {
coll_op->variable_fn_params.root_flag = true;
coll_op->variable_fn_params.root_route = NULL;
} else {
coll_op->variable_fn_params.root_flag = false;
coll_op->variable_fn_params.root_route = &topo->route_vector[root];
}
/* NTH: This was copied from the old allreduce launcher. */
if (0 < fn_idx) {
coll_op->variable_fn_params.sbuf = coll_op->variable_fn_params.rbuf;
coll_op->variable_fn_params.userbuf = coll_op->variable_fn_params.rbuf;
}
return OMPI_SUCCESS;
}
static int mca_coll_ml_allreduce_frag_progress(mca_coll_ml_collective_operation_progress_t *coll_op)
{
/* local variables */
void *buf;
size_t dt_size;
int ret, frag_len, count;
ptrdiff_t lb, extent;
Fixing broken dependency between ML/BCOLS This is hot-fix patch for the issue reported by Ralph. In future we plan to restructure ml data structure layout. Tested by Nathan. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30619. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-07 23:15:45 +04:00
mca_bcol_base_payload_buffer_desc_t *src_buffer_desc;
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_coll_ml_collective_operation_progress_t *new_op;
mca_coll_ml_module_t *ml_module = OP_ML_MODULE(coll_op);
ret = ompi_datatype_get_extent(coll_op->variable_fn_params.dtype, &lb, &extent);
if (ret < 0) {
return OMPI_ERROR;
}
dt_size = (size_t) extent;
/* Keep the pipeline filled with fragments */
while (coll_op->fragment_data.message_descriptor->n_active <
coll_op->fragment_data.message_descriptor->pipeline_depth) {
/* If an active fragment happens to have completed the collective during
* a hop into the progress engine, then don't launch a new fragment,
* instead break and return.
*/
if (coll_op->fragment_data.message_descriptor->n_bytes_scheduled
== coll_op->fragment_data.message_descriptor->n_bytes_total) {
break;
}
/* Get an ml buffer */
src_buffer_desc = mca_coll_ml_alloc_buffer(OP_ML_MODULE(coll_op));
if (NULL == src_buffer_desc) {
/* If there exist outstanding fragments, then break out
* and let an active fragment deal with this later,
* there are no buffers available.
*/
if (0 < coll_op->fragment_data.message_descriptor->n_active) {
return OMPI_SUCCESS;
}
/* It is useless to call progress from here, since
* ml progress can't be executed as result ml memsync
* call will not be completed and no memory will be
* recycled. So we put the element on the list, and we will
* progress it later when memsync will recycle some memory*/
/* The fragment is already on list and
* the we still have no ml resources
* Return busy */
if (!(coll_op->pending & REQ_OUT_OF_MEMORY)) {
coll_op->pending |= REQ_OUT_OF_MEMORY;
opal_list_append(&((OP_ML_MODULE(coll_op))->waiting_for_memory_list),
(opal_list_item_t *)coll_op);
ML_VERBOSE(10,("Out of resources %p adding to pending queue", coll_op));
} else {
ML_VERBOSE(10,("Out of resources %p", coll_op));
}
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
/* Get a new collective descriptor and initialize it */
new_op = mca_coll_ml_alloc_op_prog_single_frag_dag(ml_module,
ml_module->coll_ml_allreduce_functions[coll_op->fragment_data.current_coll_op],
coll_op->fragment_data.message_descriptor->src_user_addr,
coll_op->fragment_data.message_descriptor->dest_user_addr,
coll_op->fragment_data.message_descriptor->n_bytes_total,
coll_op->fragment_data.message_descriptor->n_bytes_scheduled);
MCA_COLL_IBOFFLOAD_SET_ML_BUFFER_INFO(new_op,
src_buffer_desc->buffer_index, src_buffer_desc);
new_op->fragment_data.current_coll_op = coll_op->fragment_data.current_coll_op;
new_op->fragment_data.message_descriptor = coll_op->fragment_data.message_descriptor;
/* set the task setup callback */
new_op->sequential_routine.seq_task_setup = mca_coll_ml_allreduce_task_setup;
/* We need this address for pointer arithmetic in memcpy */
buf = coll_op->fragment_data.message_descriptor->src_user_addr;
/* calculate the number of data types in this packet */
count = (coll_op->fragment_data.message_descriptor->n_bytes_total -
coll_op->fragment_data.message_descriptor->n_bytes_scheduled <
(size_t) OP_ML_MODULE(coll_op)->small_message_thresholds[BCOL_ALLREDUCE] ?
(coll_op->fragment_data.message_descriptor->n_bytes_total -
coll_op->fragment_data.message_descriptor->n_bytes_scheduled) / dt_size :
(size_t) coll_op->variable_fn_params.count);
/* calculate the fragment length */
frag_len = count*dt_size;
ret = ompi_datatype_copy_content_same_ddt(coll_op->variable_fn_params.dtype, count,
(char *) src_buffer_desc->data_addr, (char *) ((uintptr_t) buf + (uintptr_t)
coll_op->fragment_data.message_descriptor->n_bytes_scheduled));
if (ret < 0) {
return OMPI_ERROR;
}
/* No unpack for root */
new_op->process_fn = mca_coll_ml_allreduce_small_unpack;
/* Setup fragment specific data */
new_op->fragment_data.message_descriptor->n_bytes_scheduled += frag_len;
new_op->fragment_data.buffer_desc = src_buffer_desc;
new_op->fragment_data.fragment_size = frag_len;
(new_op->fragment_data.message_descriptor->n_active)++;
ML_SET_VARIABLE_PARAMS_BCAST(
new_op,
OP_ML_MODULE(new_op),
count,
MPI_BYTE,
src_buffer_desc,
0,
0,
frag_len,
src_buffer_desc->data_addr);
/* Fill in bcast specific arguments */
/* TBD: remove buffer_size */
new_op->variable_fn_params.buffer_size = frag_len;
new_op->variable_fn_params.count = count;
new_op->variable_fn_params.hier_factor = coll_op->variable_fn_params.hier_factor;
new_op->variable_fn_params.op = coll_op->variable_fn_params.op;
new_op->variable_fn_params.dtype = coll_op->variable_fn_params.dtype;
new_op->variable_fn_params.root = 0;
new_op->variable_fn_params.sbuf = src_buffer_desc->data_addr;
new_op->variable_fn_params.rbuf = src_buffer_desc->data_addr;
new_op->sequential_routine.current_bcol_status = SEQ_TASK_PENDING;
MCA_COLL_ML_SET_NEW_FRAG_ORDER_INFO(new_op);
ML_VERBOSE(10,("FFFF Contig + fragmentation [0-sk, 1-lk, 3-su, 4-lu] %d %d %d\n",
new_op->variable_fn_params.buffer_size,
new_op->fragment_data.fragment_size,
new_op->fragment_data.message_descriptor->n_bytes_scheduled));
/* initialize first coll */
ret = new_op->sequential_routine.seq_task_setup(new_op);
if (OMPI_SUCCESS != ret) {
ML_VERBOSE(3,("Fragment failed to initialize itself"));
return ret;
}
/* append this collective !! */
OPAL_THREAD_LOCK(&(mca_coll_ml_component.sequential_collectives_mutex));
opal_list_append(&mca_coll_ml_component.sequential_collectives,
(opal_list_item_t *)new_op);
OPAL_THREAD_UNLOCK(&(mca_coll_ml_component.sequential_collectives_mutex));
}
return OMPI_SUCCESS;
}
static inline __opal_attribute_always_inline__
int parallel_allreduce_start(void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype, struct ompi_op_t *op,
struct ompi_communicator_t *comm,
mca_coll_ml_module_t *ml_module,
ompi_request_t **req,
int small_data_allreduce,
int large_data_allreduce)
{
int ret, n_fragments = 1, frag_len,
pipeline_depth, n_dts_per_frag ;
ptrdiff_t lb, extent;
size_t pack_len, dt_size;
Fixing broken dependency between ML/BCOLS This is hot-fix patch for the issue reported by Ralph. In future we plan to restructure ml data structure layout. Tested by Nathan. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30619. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-07 23:15:45 +04:00
mca_bcol_base_payload_buffer_desc_t *src_buffer_desc;
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_coll_ml_collective_operation_progress_t *coll_op;
mca_coll_ml_component_t *cm = &mca_coll_ml_component;
bool contiguous = ompi_datatype_is_contiguous_memory_layout(dtype, count);
if (MPI_IN_PLACE == sbuf) {
sbuf = rbuf;
}
ret = ompi_datatype_get_extent(dtype, &lb, &extent);
if (ret < 0) {
return OMPI_ERROR;
}
dt_size = (size_t) extent;
pack_len = count * dt_size;
ML_VERBOSE(1,("The allreduce requested %d enable fragmentation %d ",
pack_len,
cm->enable_fragmentation));
if (pack_len <= (size_t) ml_module->small_message_thresholds[BCOL_ALLREDUCE]) {
/* The len of the message can not be larger than ML buffer size */
assert(pack_len <= ml_module->payload_block->size_buffer);
ML_VERBOSE(1,("Using small data allreduce (threshold = %d)",
ml_module->small_message_thresholds[BCOL_ALLREDUCE]));
src_buffer_desc = mca_coll_ml_alloc_buffer(ml_module);
while (OPAL_UNLIKELY(NULL == src_buffer_desc)) {
opal_progress();
src_buffer_desc = mca_coll_ml_alloc_buffer(ml_module);
}
coll_op = mca_coll_ml_alloc_op_prog_single_frag_dag(ml_module,
ml_module->coll_ml_allreduce_functions[small_data_allreduce],
sbuf, rbuf, pack_len, 0);
coll_op->variable_fn_params.rbuf = src_buffer_desc->data_addr;
coll_op->variable_fn_params.sbuf = src_buffer_desc->data_addr;
coll_op->variable_fn_params.count = count;
ret = ompi_datatype_copy_content_same_ddt(dtype, count,
(void *) (uintptr_t) src_buffer_desc->data_addr, (char *) sbuf);
if (ret < 0){
return OMPI_ERROR;
}
/* unpack function */
coll_op->process_fn = mca_coll_ml_allreduce_small_unpack;
} else if (cm->enable_fragmentation || !contiguous) {
ML_VERBOSE(1,("Using Fragmented Allreduce"));
/* fragment the data */
/* check for retarded application programming decisions */
if (dt_size > (size_t) ml_module->small_message_thresholds[BCOL_ALLREDUCE]) {
ML_ERROR(("Sorry, but we don't support datatypes that large"));
return OMPI_ERROR;
}
/* calculate the number of data types that can fit per ml-buffer */
n_dts_per_frag = ml_module->small_message_thresholds[BCOL_ALLREDUCE] / dt_size;
/* calculate the number of fragments */
n_fragments = (count + n_dts_per_frag - 1) / n_dts_per_frag; /* round up */
/* calculate the actual pipeline depth */
pipeline_depth = n_fragments < cm->pipeline_depth ? n_fragments : cm->pipeline_depth;
/* calculate the fragment size */
frag_len = n_dts_per_frag * dt_size;
/* allocate an ml buffer */
src_buffer_desc = mca_coll_ml_alloc_buffer(ml_module);
while (NULL == src_buffer_desc) {
opal_progress();
src_buffer_desc = mca_coll_ml_alloc_buffer(ml_module);
}
coll_op = mca_coll_ml_alloc_op_prog_single_frag_dag(ml_module,
ml_module->coll_ml_allreduce_functions[small_data_allreduce],
sbuf, rbuf, pack_len, 0 /* offset for first pack */);
/* task setup callback function */
coll_op->sequential_routine.seq_task_setup = mca_coll_ml_allreduce_task_setup;
coll_op->process_fn = mca_coll_ml_allreduce_small_unpack;
coll_op->variable_fn_params.sbuf = (void *) src_buffer_desc->data_addr;
coll_op->variable_fn_params.rbuf = (void *) src_buffer_desc->data_addr;
coll_op->fragment_data.message_descriptor->n_active = 1;
coll_op->full_message.n_bytes_scheduled = frag_len;
coll_op->full_message.fragment_launcher = mca_coll_ml_allreduce_frag_progress;
coll_op->full_message.pipeline_depth = pipeline_depth;
coll_op->fragment_data.current_coll_op = small_data_allreduce;
coll_op->fragment_data.fragment_size = frag_len;
coll_op->variable_fn_params.count = n_dts_per_frag; /* seems fishy */
coll_op->variable_fn_params.buffer_size = frag_len;
/* copy into the ml-buffer */
ret = ompi_datatype_copy_content_same_ddt(dtype, n_dts_per_frag,
(char *) src_buffer_desc->data_addr, (char *) sbuf);
if (ret < 0) {
return OMPI_ERROR;
}
} else {
ML_VERBOSE(1,("Using zero-copy ptp allreduce"));
coll_op = mca_coll_ml_alloc_op_prog_single_frag_dag(ml_module,
ml_module->coll_ml_allreduce_functions[large_data_allreduce],
sbuf, rbuf, pack_len, 0);
coll_op->variable_fn_params.userbuf =
coll_op->variable_fn_params.sbuf = sbuf;
coll_op->variable_fn_params.rbuf = rbuf;
/* The ML buffer is used for testing. Later, when we
* switch to use knem/mmap/portals this should be replaced
* appropriately
*/
src_buffer_desc = mca_coll_ml_alloc_buffer(ml_module);
while (NULL == src_buffer_desc) {
opal_progress();
src_buffer_desc = mca_coll_ml_alloc_buffer(ml_module);
}
coll_op->variable_fn_params.count = count;
}
MCA_COLL_IBOFFLOAD_SET_ML_BUFFER_INFO(coll_op, src_buffer_desc->buffer_index,
src_buffer_desc);
/* set the offset */
coll_op->variable_fn_params.sbuf_offset = 0;
coll_op->variable_fn_params.rbuf_offset = 0;
/* Fill in the function arguments */
coll_op->variable_fn_params.sequence_num =
coll/ml: fix some warnings and the spelling of indices This commit fixes one warning that should have caused coll/ml to segfault on reduce. The fix should be correct but we will continue to investigate. cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30477. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-01-29 22:44:21 +04:00
OPAL_THREAD_ADD32(&(ml_module->collective_sequence_num), 1);
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
coll_op->sequential_routine.current_active_bcol_fn = 0;
coll_op->variable_fn_params.dtype = dtype;
coll_op->variable_fn_params.op = op;
coll_op->variable_fn_params.root = 0;
coll_op->sequential_routine.seq_task_setup = mca_coll_ml_allreduce_task_setup; /* invoked after each level in sequential
* progress call
*/
MCA_COLL_ML_SET_ORDER_INFO(coll_op, n_fragments);
ret = mca_coll_ml_launch_sequential_collective (coll_op);
if (ret != OMPI_SUCCESS) {
ML_VERBOSE(10, ("Failed to launch"));
return ret;
}
*req = &coll_op->full_message.super;
return OMPI_SUCCESS;
}
int mca_coll_ml_allreduce(void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype, struct ompi_op_t *op,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_ml_module_t *ml_module = (mca_coll_ml_module_t*)module;
ompi_request_t *req;
int ret;
if (OPAL_UNLIKELY(!ompi_op_is_commute(op))) {
/* coll/ml does not handle non-communative operations at this time. fallback
* on another collective module */
return ml_module->fallback.coll_allreduce (sbuf, rbuf, count, dtype, op, comm,
ml_module->fallback.coll_allreduce_module);
}
ret = parallel_allreduce_start(sbuf, rbuf, count, dtype, op, comm,
(mca_coll_ml_module_t *) module, &req,
ML_SMALL_DATA_ALLREDUCE,
ML_LARGE_DATA_ALLREDUCE);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
ML_ERROR(("Failed to launch"));
return ret;
}
ompi_request_wait_completion(req);
ompi_request_free(&req);
ML_VERBOSE(10, ("Blocking NB allreduce is done"));
return OMPI_SUCCESS;
}
int mca_coll_ml_allreduce_nb(void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype, struct ompi_op_t *op,
struct ompi_communicator_t *comm,
ompi_request_t **req,
mca_coll_base_module_t *module)
{
mca_coll_ml_module_t *ml_module = (mca_coll_ml_module_t*)module;
int ret;
if (OPAL_UNLIKELY(!ompi_op_is_commute(op))) {
/* coll/ml does not handle non-communative operations at this time. fallback
* on another collective module */
return ml_module->fallback.coll_iallreduce (sbuf, rbuf, count, dtype, op, comm, req,
ml_module->fallback.coll_iallreduce_module);
}
ret = parallel_allreduce_start(sbuf, rbuf, count, dtype, op, comm,
(mca_coll_ml_module_t *) module, req,
ML_SMALL_DATA_ALLREDUCE,
ML_LARGE_DATA_ALLREDUCE);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
ML_ERROR(("Failed to launch"));
return ret;
}
ML_VERBOSE(10, ("Blocking NB allreduce is done"));
return OMPI_SUCCESS;
}
int mca_coll_ml_allreduce_dispatch(void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype, struct ompi_op_t *op,
struct ompi_communicator_t *comm, mca_coll_base_module_t *module)
{
int rc;
bool use_extra_topo;
ompi_request_t *req;
mca_coll_ml_module_t *ml_module = (mca_coll_ml_module_t *) module;
use_extra_topo = (count > 1) ?
!ml_module->allreduce_matrix[op->op_type][dtype->id][BCOL_MULTI_ELEM_TYPE] :
!ml_module->allreduce_matrix[op->op_type][dtype->id][BCOL_SINGLE_ELEM_TYPE];
if (use_extra_topo) {
rc = parallel_allreduce_start(sbuf, rbuf, count, dtype,
op, comm, ml_module, &req,
ML_SMALL_DATA_EXTRA_TOPO_ALLREDUCE,
ML_LARGE_DATA_EXTRA_TOPO_ALLREDUCE);
} else {
rc = parallel_allreduce_start(sbuf, rbuf, count, dtype,
op, comm, ml_module, &req,
ML_SMALL_DATA_ALLREDUCE,
ML_LARGE_DATA_ALLREDUCE);
}
if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
ML_ERROR(("Failed to launch"));
return rc;
}
ompi_request_wait_completion(req);
ompi_request_free(&req);
return OMPI_SUCCESS;
}
int mca_coll_ml_allreduce_dispatch_nb(void *sbuf, void *rbuf, int count,
ompi_datatype_t *dtype, ompi_op_t *op,
ompi_communicator_t *comm,
ompi_request_t **req,
mca_coll_base_module_t *module)
{
int rc;
bool use_extra_topo;
mca_coll_ml_module_t *ml_module = (mca_coll_ml_module_t *) module;
use_extra_topo = (count > 1) ?
!ml_module->allreduce_matrix[op->op_type][dtype->id][BCOL_MULTI_ELEM_TYPE] :
!ml_module->allreduce_matrix[op->op_type][dtype->id][BCOL_SINGLE_ELEM_TYPE];
if (use_extra_topo) {
rc = parallel_allreduce_start(sbuf, rbuf, count, dtype,
op, comm, ml_module, req,
ML_SMALL_DATA_EXTRA_TOPO_ALLREDUCE,
ML_LARGE_DATA_EXTRA_TOPO_ALLREDUCE);
} else {
rc = parallel_allreduce_start(sbuf, rbuf, count, dtype,
op, comm, ml_module, req,
ML_SMALL_DATA_ALLREDUCE,
ML_LARGE_DATA_ALLREDUCE);
}
if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
ML_ERROR(("Failed to launch"));
return rc;
}
return OMPI_SUCCESS;
}
Копировать постоянную ссылку