ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
f45914db9b
openmpi/ompi/mca/coll/ml/coll_ml_allgather.c

632 строки
26 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.
coll/ml: cleanup some merge related errors cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30366. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-01-22 20:48:09 +04:00
* Copyright (c) 2013-2014 Los Alamos National Security, LLC. All rights
coll/ml: add support for blocking and non-blocking allreduce, reduce, and allgather. The new collectives provide a signifigant performance increase over tuned for small and medium messages. We are initially setting the priority lower than tuned until this has had some time to soak in the trunk. Please set coll_ml_priority to 90 for MTT runs. Credit for this work goes to Manjunath Gorentla Venkata (ORNL), Pavel Shamis (ORNL), and Nathan Hjelm (LANL). Commit details (for reference): Import ORNL's collectives for MPI_Allreduce, MPI_Reduce, and MPI_Allgather. We need to take the basesmuma header into account when calculating the ptpcoll small message thresholds. Add a define to bcol.h indicating the maximum header size so we can take the header into account while not making ptpcoll dependent on information from basesmuma. This resolves an issue with allreduce where ptpcoll overwrites the header of the next buffer in the basesmuma bank. Fix reduce and make a sequential collective launcher in coll_ml_inlines.h The root calculation for reduce was wrong for any root != 0. There are four possibilities for the root: - The root is not the current process but is in the current hierarchy. In this case the root is the index of the global root as specified in the root vector. - The root is not the current process and is not in the next level of the hierarchy. In this case 0 must be the local root since this process will never communicate with the real root. - The root is not the current process but will be in next level of the hierarchy. In this case the current process must be the root. - I am the root. The root is my index. Tested with IMB which rotates the root on every call to MPI_Reduce. Consider IMB the reproducer for the issue this commit solves. Make the bcast algorithm decision an enumerated variable Resolve various asset failures when destructing coll ml requests. Two issues: - Always reset the request to be invalid before returning it to the free list. This will avoid an asset in ompi_request_t's destructor. OMPI_REQUEST_FINI does this (and also releases the fortran handle index). - Never explicitly construct or destruct the superclass of an opal object. This screws up the class function tables and will cause either an assert failure or a segmentation fault when destructing coll ml requests. Cleanup allgather. I removed the duplicate non-blocking and blocking functions and modeled the cleanup after what I found in allreduce. Also cleaned up the code somewhat. Don't bother copying from the send to the recieve buffer in bcol_basesmuma_allreduce_intra_fanin_fanout if the pointers are the same. The eliminates a warning about memcpy and aliasing and avoids an unnecessary call to memcpy. Alwasy call CHECK_AND_RELEASE on memsync collectives. There was a call to OBJ_RELEASE on the collective communicator but because CHECK_AND_RECYLCE was never called there was not matching call to OBJ_RELEASE. This caused coll ml to leak communicators. Make allreduce use the sequential collective launcher in coll_ml_inlines.h Just launch the next collective in the component progress. I am a little unsure about this patch. There appears to be some sort of race between collectives that causes buffer exhaustion in some cases (IMB Allreduce is a reproducer). Changing progress to only launch the next bcol seems to resolve the issue but might not be the best fix. Note that I see little-no performance penalty for this change. Fix allreduce when there are extra sources. There was an issue with the buffer offset calculation when there are extra sources. In the case of extra sources == 1 the offset was set to buffer_size (just past the header of the next buffer). I adjusted the buffer size to take into accoun the maximum header size (see the earlier commit that added this) and simplified the offset calculation. Make reduce/allreduce non-blocking. This is required for MPI_Comm_idup to work correctly. This has been tested with various layouts using the ibm testsuite and imb and appears to have the same performance as the old blocking version. Fix allgather for non-contiguous layouts and simplify parsing the topology. Some things in this patch: - There were several comments to the effect that level 0 of the hierarchy MUST contain all of the ranks. At least one function made this assumption but it was not true. I changed the sbgp components and the coll ml initization code to enforce this requirement. - Ensure that hierarchy level 0 has the ranks in the correct scatter gather order. This removes the need for a separate sort list and fixes the offset calculation for allgather. - There were several passes over the hierarchy to determine properties of the hierarchy. I eliminated these extra passes and the memory allocation associated with them and calculate the tree properties on the fly. The same DFS recursion also handles the re-order of level 0. All these changes have been verified with MPI_Allreduce, MPI_Reduce, and MPI_Allgather. All functions now pass all IBM/Open MPI, and IMB tests. coll/ml: correct pointer usage for MPI_BOTTOM Since contiguous datatypes are copied via memcpy (bypassing the convertor) we need to adjust for the lb of the datatype. This corrects problems found testing code that uses MPI_BOTTOM (NULL) as the send pointer. Add fallback collectives for allreduce and reduce. cmr=v1.7.5:reviewer=pasha This commit was SVN r30363.
2014-01-22 19:39:19 +04:00
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/** @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_allgather_small_unpack_data(mca_coll_ml_collective_operation_progress_t *coll_op)
{
bool rcontig = coll_op->full_message.recv_data_continguous;
int n_ranks_in_comm = ompi_comm_size(OP_ML_MODULE(coll_op)->comm);
void *dest = (void *)((uintptr_t)coll_op->full_message.dest_user_addr +
(uintptr_t)coll_op->full_message.n_bytes_delivered);
void *src = (void *)((uintptr_t)coll_op->fragment_data.buffer_desc->data_addr +
(size_t)coll_op->variable_fn_params.rbuf_offset);
if (rcontig) {
memcpy(dest, src, n_ranks_in_comm * coll_op->full_message.n_bytes_scheduled);
} else {
mca_coll_ml_convertor_unpack(src, n_ranks_in_comm * coll_op->full_message.n_bytes_scheduled,
&coll_op->fragment_data.message_descriptor->recv_convertor);
}
return OMPI_SUCCESS;
}
static inline void copy_data (mca_coll_ml_collective_operation_progress_t *coll_op, rank_properties_t *rank_props, int soffset) {
bool rcontig = coll_op->fragment_data.message_descriptor->recv_data_continguous;
size_t total_bytes = coll_op->fragment_data.message_descriptor->n_bytes_total;
size_t pack_len = coll_op->fragment_data.fragment_size;
int doffset = rank_props->rank;
void *dest, *src;
src = (void *) ((uintptr_t)coll_op->fragment_data.buffer_desc->data_addr +
(size_t)coll_op->variable_fn_params.rbuf_offset + soffset * pack_len);
if (rcontig) {
dest = (void *) ((uintptr_t) coll_op->full_message.dest_user_addr +
(uintptr_t) coll_op->fragment_data.offset_into_user_buffer +
doffset * total_bytes);
memcpy(dest, src, pack_len);
} else {
size_t position;
opal_convertor_t *recv_convertor =
&coll_op->fragment_data.message_descriptor->recv_convertor;
position = (size_t) coll_op->fragment_data.offset_into_user_buffer +
doffset * total_bytes;
opal_convertor_set_position(recv_convertor, &position);
mca_coll_ml_convertor_unpack(src, pack_len, recv_convertor);
}
}
static int mca_coll_ml_allgather_noncontiguous_unpack_data(mca_coll_ml_collective_operation_progress_t *coll_op)
{
int i, j, n_level_one_sbgps;
size_t soffset;
mca_coll_ml_topology_t *topo_info = coll_op->coll_schedule->topo_info;
sub_group_params_t *array_of_all_subgroup_ranks = topo_info->array_of_all_subgroups;
n_level_one_sbgps = array_of_all_subgroup_ranks->level_one_index;
for (i = 0 ; i < n_level_one_sbgps; i++) {
/* determine where in the source buffer the data can be found */
soffset = array_of_all_subgroup_ranks[i].index_of_first_element;
for (j = 0 ; j < array_of_all_subgroup_ranks[i].n_ranks; j++, ++soffset) {
copy_data (coll_op, array_of_all_subgroup_ranks[i].rank_data + j, soffset);
}
}
return OMPI_SUCCESS;
}
/* Allgather dependencies seem easy, everyone needs to work from the "bottom up".
* Following Pasha, I too will put the simplest dependencies graph and change it later
* when we add hierarchy. Basically, allgather has the same dependency profile as the
* sequential broadcast except that there is only a single ordering of tasks.
*/
static int mca_coll_ml_allgather_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 allgather, 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];
}
return OMPI_SUCCESS;
}
static int mca_coll_ml_allgather_frag_progress(mca_coll_ml_collective_operation_progress_t *coll_op)
{
/* local variables */
int ret;
size_t frag_len, dt_size;
void *buf;
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);
bool scontig = coll_op->fragment_data.message_descriptor->send_data_continguous;
ompi_datatype_type_size(coll_op->variable_fn_params.dtype, &dt_size);
/* 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(ml_module);
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;
} else {
/* The fragment is already on list and
* the we still have no ml resources
* Return busy */
if (coll_op->pending & REQ_OUT_OF_MEMORY) {
ML_VERBOSE(10,("Out of resources %p", coll_op));
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
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));
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_allgather_functions[ML_SMALL_DATA_ALLGATHER],
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);
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_allgather_task_setup;
/*
MCA_COLL_IBOFFLOAD_SET_ML_BUFFER_INFO(new_op,
src_buffer_desc->buffer_index, src_buffer_desc);
*/
/* We need this address for pointer arithmetic in memcpy */
buf = coll_op->fragment_data.message_descriptor->src_user_addr;
if (!scontig) {
frag_len = ml_module->small_message_thresholds[BCOL_ALLGATHER];
mca_coll_ml_convertor_get_send_frag_size(
ml_module, &frag_len,
coll_op->fragment_data.message_descriptor);
mca_coll_ml_convertor_pack(
(void *) ((uintptr_t) src_buffer_desc->data_addr +
frag_len * coll_op->coll_schedule->topo_info->hier_layout_info[0].offset +
frag_len * coll_op->coll_schedule->topo_info->hier_layout_info[0].level_one_index),
frag_len, &coll_op->fragment_data.message_descriptor->send_convertor);
} else {
/* calculate new frag length, there are some issues here */
frag_len = (coll_op->fragment_data.message_descriptor->n_bytes_total -
coll_op->fragment_data.message_descriptor->n_bytes_scheduled <
coll_op->fragment_data.fragment_size ?
coll_op->fragment_data.message_descriptor->n_bytes_total -
coll_op->fragment_data.message_descriptor->n_bytes_scheduled :
coll_op->fragment_data.fragment_size);
/* everybody copies in, based on the new values */
memcpy((void *) ((uintptr_t)src_buffer_desc->data_addr +
frag_len * new_op->coll_schedule->topo_info->hier_layout_info[0].offset +
frag_len * new_op->coll_schedule->topo_info->hier_layout_info[0].level_one_index),
(void *) ((uintptr_t) buf + (uintptr_t)
coll_op->fragment_data.message_descriptor->n_bytes_scheduled), frag_len);
}
new_op->variable_fn_params.sbuf = (void *) src_buffer_desc->data_addr;
new_op->variable_fn_params.rbuf = (void *) src_buffer_desc->data_addr;
/* update the number of bytes scheduled */
new_op->fragment_data.message_descriptor->n_bytes_scheduled += frag_len;
/* everyone needs an unpack function */
new_op->process_fn = mca_coll_ml_allgather_noncontiguous_unpack_data;
new_op->fragment_data.fragment_size = frag_len;
new_op->fragment_data.buffer_desc = src_buffer_desc;
/* Setup fragment specific data */
++(new_op->fragment_data.message_descriptor->n_active);
ML_VERBOSE(10, ("Start more, My index %d ",
new_op->fragment_data.buffer_desc->buffer_index));
/* this is a bit buggy */
ML_SET_VARIABLE_PARAMS_BCAST(
new_op,
OP_ML_MODULE(new_op),
frag_len /* yes, we have consistent units, so this makes sense */,
MPI_BYTE /* we fragment according to buffer size
* we don't reduce the data thus we needn't
* keep "whole" datatypes, we may freely
* fragment without regard for multiples
* of any specific datatype
*/,
src_buffer_desc,
0,
0,
frag_len,
src_buffer_desc->data_addr);
/* 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;
}
new_op->variable_fn_params.buffer_size = frag_len;
new_op->variable_fn_params.hier_factor = coll_op->variable_fn_params.hier_factor;
new_op->variable_fn_params.root = 0;
MCA_COLL_ML_SET_NEW_FRAG_ORDER_INFO(new_op);
/* 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 mca_coll_ml_allgather_start (void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void* rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module,
ompi_request_t **req)
{
size_t pack_len, sdt_size;
int ret, n_fragments = 1, comm_size;
mca_coll_ml_topology_t *topo_info;
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_component_t *cm = &mca_coll_ml_component;
mca_coll_ml_collective_operation_progress_t *coll_op;
mca_coll_ml_module_t *ml_module = (mca_coll_ml_module_t *) module;
ptrdiff_t lb, extent;
bool scontig, rcontig, in_place = false;
/* check for in place setting */
if (MPI_IN_PLACE == sbuf) {
in_place = true;
sdtype = rdtype;
scount = rcount;
}
/* scontig could be != to rcontig */
scontig = ompi_datatype_is_contiguous_memory_layout(sdtype, scount);
rcontig = ompi_datatype_is_contiguous_memory_layout(rdtype, rcount);
comm_size = ompi_comm_size(comm);
ML_VERBOSE(10, ("Starting allgather"));
assert(NULL != sdtype);
/* Calculate size of the data,
* at this stage, only contiguous data is supported */
/* this is valid for allagther */
ompi_datatype_type_size(sdtype, &sdt_size);
pack_len = scount * sdt_size;
if (in_place) {
coll/ml: cleanup some merge related errors cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30366. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-01-22 20:48:09 +04:00
sbuf = (char *) rbuf + ompi_comm_rank(comm) * pack_len;
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
}
/* Allocate collective schedule and pack message */
/* this is the total ending message size that will need to fit in the ml-buffer */
if (pack_len <= (size_t) ml_module->small_message_thresholds[BCOL_ALLGATHER]) {
/* The len of the message can not be larger than ML buffer size */
ML_VERBOSE(10, ("Single frag %d %d %d", pack_len, comm_size, ml_module->payload_block->size_buffer));
assert(pack_len * comm_size <= ml_module->payload_block->size_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);
}
/* change 1 */
coll_op = mca_coll_ml_alloc_op_prog_single_frag_dag(ml_module,
ml_module->coll_ml_allgather_functions[ML_SMALL_DATA_ALLGATHER],
sbuf, rbuf, pack_len, 0 /* offset for first pack */);
MCA_COLL_IBOFFLOAD_SET_ML_BUFFER_INFO(coll_op,
src_buffer_desc->buffer_index, src_buffer_desc);
coll_op->fragment_data.current_coll_op = ML_SMALL_DATA_ALLGATHER;
/* task setup callback function */
coll_op->sequential_routine.seq_task_setup = mca_coll_ml_allgather_task_setup;
/* change 2 */
if (!scontig) {
coll_op->full_message.n_bytes_scheduled =
mca_coll_ml_convertor_prepare(sdtype, scount, sbuf,
&coll_op->full_message.send_convertor, MCA_COLL_ML_NET_STREAM_SEND);
mca_coll_ml_convertor_pack(
(void *) ((uintptr_t) src_buffer_desc->data_addr + pack_len *
(coll_op->coll_schedule->topo_info->hier_layout_info[0].offset +
coll_op->coll_schedule->topo_info->hier_layout_info[0].level_one_index)),
pack_len, &coll_op->full_message.send_convertor);
} else {
/* change 3 */
memcpy((void *)((uintptr_t) src_buffer_desc->data_addr + pack_len *
(coll_op->coll_schedule->topo_info->hier_layout_info[0].offset +
coll_op->coll_schedule->topo_info->hier_layout_info[0].level_one_index)),
sbuf, pack_len);
coll_op->full_message.n_bytes_scheduled = pack_len;
}
if (!rcontig) {
mca_coll_ml_convertor_prepare(rdtype, rcount * comm_size, rbuf,
&coll_op->full_message.recv_convertor, MCA_COLL_ML_NET_STREAM_RECV);
}
if (coll_op->coll_schedule->topo_info->ranks_contiguous) {
coll_op->process_fn = mca_coll_ml_allgather_small_unpack_data;
} else {
coll_op->process_fn = mca_coll_ml_allgather_noncontiguous_unpack_data;
}
/* whole ml-buffer is used to send AND receive */
coll_op->variable_fn_params.sbuf = (void *) src_buffer_desc->data_addr;
coll_op->variable_fn_params.rbuf = (void *) src_buffer_desc->data_addr;
/* we can set the initial offset here */
coll_op->variable_fn_params.sbuf_offset = 0;
coll_op->variable_fn_params.rbuf_offset = 0;
coll_op->variable_fn_params.count = scount;
coll_op->fragment_data.fragment_size =
coll_op->full_message.n_bytes_scheduled;
/* For small CINCO, we may use the native datatype */
coll_op->variable_fn_params.dtype = sdtype;
coll_op->variable_fn_params.buffer_size = pack_len;
coll_op->variable_fn_params.root = 0;
} else if (cm->enable_fragmentation || pack_len * comm_size < (1 << 20)) {
/* calculate the number of fragments and the size of each frag */
size_t n_dts_per_frag, frag_len;
int pipeline_depth = mca_coll_ml_component.pipeline_depth;
/* Calculate the number of fragments required for this message careful watch the integer division !*/
frag_len = (pack_len <= (size_t) ml_module->small_message_thresholds[BCOL_ALLGATHER] ?
pack_len : (size_t) ml_module->small_message_thresholds[BCOL_ALLGATHER]);
n_dts_per_frag = frag_len / sdt_size;
n_fragments = (pack_len + sdt_size * n_dts_per_frag - 1) / (sdt_size * n_dts_per_frag);
pipeline_depth = (n_fragments < pipeline_depth ? n_fragments : pipeline_depth);
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);
}
/* change 4 */
coll_op = mca_coll_ml_alloc_op_prog_single_frag_dag(ml_module,
ml_module->coll_ml_allgather_functions[ML_SMALL_DATA_ALLGATHER],
sbuf, rbuf, pack_len,
0 /* offset for first pack */);
MCA_COLL_IBOFFLOAD_SET_ML_BUFFER_INFO(coll_op,
src_buffer_desc->buffer_index, src_buffer_desc);
topo_info = coll_op->coll_schedule->topo_info;
/* task setup callback function */
coll_op->sequential_routine.seq_task_setup = mca_coll_ml_allgather_task_setup;
if (!scontig) {
coll_op->full_message.send_converter_bytes_packed =
mca_coll_ml_convertor_prepare(
sdtype, scount, NULL,
&coll_op->full_message.dummy_convertor,
MCA_COLL_ML_NET_STREAM_SEND);
coll_op->full_message.dummy_conv_position = 0;
mca_coll_ml_convertor_get_send_frag_size(
ml_module, &frag_len,
&coll_op->full_message);
/* change 5 */
mca_coll_ml_convertor_prepare(sdtype, scount, sbuf,
&coll_op->full_message.send_convertor, MCA_COLL_ML_NET_STREAM_SEND);
mca_coll_ml_convertor_pack(
(void *) ((uintptr_t) src_buffer_desc->data_addr + frag_len *
coll/ml: fix issues identified by the clang static analyser and fix a segmentation fault in the reduce cleanup Some of the changes address false warnings produced by scan-build. I added asserts and changed some malloc calls to calloc to silence these warnings. The was one issue in cleanup for reduce since the component_functions member is changed by the allreduce call. There may be other issues with how this code works but releasing the allocated component_functions after setting up the static functions addresses the primary issue (SIGSEGV). cmr=v1.8.1:reviewer=manjugv This commit was SVN r31417.
2014-04-17 02:43:35 +04:00
(topo_info->hier_layout_info[0].offset +
topo_info->hier_layout_info[0].level_one_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
frag_len, &coll_op->full_message.send_convertor);
} else {
/* change 6 */
memcpy((void *)((uintptr_t)src_buffer_desc->data_addr + frag_len *
coll/ml: fix issues identified by the clang static analyser and fix a segmentation fault in the reduce cleanup Some of the changes address false warnings produced by scan-build. I added asserts and changed some malloc calls to calloc to silence these warnings. The was one issue in cleanup for reduce since the component_functions member is changed by the allreduce call. There may be other issues with how this code works but releasing the allocated component_functions after setting up the static functions addresses the primary issue (SIGSEGV). cmr=v1.8.1:reviewer=manjugv This commit was SVN r31417.
2014-04-17 02:43:35 +04:00
(topo_info->hier_layout_info[0].offset +
topo_info->hier_layout_info[0].level_one_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
sbuf, frag_len);
}
if (!rcontig) {
mca_coll_ml_convertor_prepare(rdtype, rcount * comm_size, rbuf,
&coll_op->full_message.recv_convertor, MCA_COLL_ML_NET_STREAM_RECV);
}
coll_op->process_fn = mca_coll_ml_allgather_noncontiguous_unpack_data;
/* hopefully this doesn't royaly screw things up idea behind this is the
* whole ml-buffer is used to send and receive
*/
coll_op->variable_fn_params.sbuf = (void *) src_buffer_desc->data_addr;
coll_op->variable_fn_params.rbuf = (void *) src_buffer_desc->data_addr;
/* we can set the initial offset here */
coll_op->variable_fn_params.sbuf_offset = 0;
coll_op->variable_fn_params.rbuf_offset = 0;
coll_op->fragment_data.buffer_desc = src_buffer_desc;
coll_op->fragment_data.fragment_size = frag_len;
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_allgather_frag_progress;
coll_op->full_message.pipeline_depth = pipeline_depth;
coll_op->fragment_data.current_coll_op = ML_SMALL_DATA_ALLGATHER;
/* remember this is different for frags !! Caused data corruption when
* not properly set. Need to be sure you have consistent units.
*/
coll_op->variable_fn_params.count = frag_len;
coll_op->variable_fn_params.dtype = MPI_BYTE; /* for fragmented data, we work in
* units of bytes. This means that
* all of our arithmetic is done
* in terms of bytes
*/
coll_op->variable_fn_params.root = 0;
coll_op->variable_fn_params.frag_size = frag_len;
coll_op->variable_fn_params.buffer_size = frag_len;
} else {
/* change 7 */
ML_VERBOSE(10, ("ML_ALLGATHER_LARGE_DATA_KNOWN case."));
coll_op = mca_coll_ml_alloc_op_prog_single_frag_dag(ml_module,
ml_module->coll_ml_allgather_functions[ML_LARGE_DATA_ALLGATHER],
sbuf, rbuf, pack_len, 0 /* offset for first pack */);
topo_info = coll_op->coll_schedule->topo_info;
if (MCA_BCOL_BASE_NO_ML_BUFFER_FOR_LARGE_MSG & topo_info->all_bcols_mode) {
MCA_COLL_IBOFFLOAD_SET_ML_BUFFER_INFO(coll_op, MCA_COLL_ML_NO_BUFFER, NULL);
} else {
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);
}
MCA_COLL_IBOFFLOAD_SET_ML_BUFFER_INFO(coll_op, src_buffer_desc->buffer_index, src_buffer_desc);
}
/* not sure if I really need this here */
coll_op->sequential_routine.seq_task_setup = mca_coll_ml_allgather_task_setup;
coll_op->process_fn = NULL;
/* probably the most important piece */
coll_op->variable_fn_params.sbuf = sbuf;
coll_op->variable_fn_params.rbuf = rbuf;
coll_op->variable_fn_params.sbuf_offset = 0;
coll_op->variable_fn_params.rbuf_offset = 0;
coll_op->variable_fn_params.count = scount;
coll_op->variable_fn_params.dtype = sdtype;/* for zero copy, we want the
* native datatype and actual count
*/
coll_op->variable_fn_params.root = 0;
/* you still need to copy in your own data into the rbuf */
/* don't need to do this if you have in place data */
if (!in_place) {
memcpy((char *) rbuf + ompi_comm_rank(comm) * pack_len, sbuf, pack_len);
}
}
coll_op->full_message.send_count = scount;
coll_op->full_message.recv_count = rcount;
coll_op->full_message.send_data_continguous = scontig;
coll_op->full_message.recv_data_continguous = rcontig;
ompi_datatype_get_extent(sdtype, &lb, &extent);
coll_op->full_message.send_extent = (size_t) extent;
ompi_datatype_get_extent(rdtype, &lb, &extent);
coll_op->full_message.recv_extent = (size_t) extent;
/* 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->variable_fn_params.hier_factor = comm_size;
MCA_COLL_ML_SET_ORDER_INFO(coll_op, n_fragments);
ret = mca_coll_ml_launch_sequential_collective (coll_op);
if (OMPI_SUCCESS != ret) {
ML_VERBOSE(10, ("Failed to launch"));
return ret;
}
*req = &coll_op->full_message.super;
return OMPI_SUCCESS;
}
int mca_coll_ml_allgather(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void* rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
ompi_request_t *req;
int ret;
ML_VERBOSE(10, ("Starting blocking allgather"));
ret = mca_coll_ml_allgather_start (sbuf, scount, sdtype,
rbuf, rcount, rdtype,
comm, module, &req);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
ret = ompi_request_wait (&req, MPI_STATUS_IGNORE);
ML_VERBOSE(10, ("Blocking allgather is complete"));
return ret;
}
int mca_coll_ml_allgather_nb(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void* rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
ompi_request_t **req,
mca_coll_base_module_t *module)
{
int ret;
ML_VERBOSE(10, ("Starting non-blocking allgather"));
ret = mca_coll_ml_allgather_start (sbuf, scount, sdtype,
rbuf, rcount, rdtype,
comm, module, req);
if (OPAL_UNLIKELY(OMPI_SUCCESS != ret)) {
return ret;
}
ML_VERBOSE(10, ("Non-blocking allgather started"));
return ret;
}
Копировать постоянную ссылку