ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
1a021b8f2d
openmpi/ompi/mca/bcol/ptpcoll/bcol_ptpcoll_module.c
Nathan Hjelm 1a021b8f2d 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 15:39:19 +00:00

721 строка
26 KiB
C
Исходник Ответственный История

/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2009-2013 Oak Ridge National Laboratory. All rights reserved.
* Copyright (c) 2009-2012 Mellanox Technologies. All rights reserved.
* Copyright (c) 2012-2013 Los Alamos National Security, LLC. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
/**
* @file
*
*/
#include "ompi_config.h"
#include <unistd.h>
#include <sys/types.h>
#include <sys/mman.h>
#include <fcntl.h>
#include <errno.h>
#include "ompi/constants.h"
#include "ompi/communicator/communicator.h"
#include "ompi/mca/bcol/bcol.h"
#include "opal/util/show_help.h"
#include "ompi/mca/bcol/base/base.h"
#include "ompi/mca/pml/pml.h" /* need this for the max tag size */
#include "ompi/mca/coll/ml/coll_ml.h"
#include "ompi/mca/coll/ml/coll_ml_allocation.h"
#include "bcol_ptpcoll.h"
#include "bcol_ptpcoll_utils.h"
#include "bcol_ptpcoll_bcast.h"
#include "bcol_ptpcoll_allreduce.h"
#include "bcol_ptpcoll_reduce.h"
#define BCOL_PTP_CACHE_LINE_SIZE 128
/*
* Local functions
*/
static int alloc_allreduce_offsets_array(mca_bcol_ptpcoll_module_t *ptpcoll_module)
{
int rc = OMPI_SUCCESS, i = 0;
netpatterns_k_exchange_node_t *k_node = &ptpcoll_module->knomial_exchange_tree;
int n_exchanges = k_node->n_exchanges;
/* Precalculate the allreduce offsets */
if (0 < k_node->n_exchanges) {
ptpcoll_module->allgather_offsets = (int **)malloc(n_exchanges * sizeof(int*));
if (!ptpcoll_module->allgather_offsets) {
rc = OMPI_ERROR;
return rc;
}
for (i=0; i < n_exchanges ; i++) {
ptpcoll_module->allgather_offsets[i] = (int *)malloc (sizeof(int) * NOFFSETS);
if (!ptpcoll_module->allgather_offsets[i]){
rc = OMPI_ERROR;
return rc;
}
}
}
return rc;
}
static int free_allreduce_offsets_array(mca_bcol_ptpcoll_module_t *ptpcoll_module)
{
int rc = OMPI_SUCCESS, i = 0;
netpatterns_k_exchange_node_t *k_node = &ptpcoll_module->knomial_exchange_tree;
int n_exchanges = k_node->n_exchanges;
if (ptpcoll_module->allgather_offsets) {
for (i=0; i < n_exchanges; i++) {
free (ptpcoll_module->allgather_offsets[i]);
}
}
free(ptpcoll_module->allgather_offsets);
return rc;
}
static void
mca_bcol_ptpcoll_module_construct(mca_bcol_ptpcoll_module_t *ptpcoll_module)
{
uint64_t i;
/* Pointer to component */
ptpcoll_module->super.bcol_component = (mca_bcol_base_component_t *) &mca_bcol_ptpcoll_component;
ptpcoll_module->super.list_n_connected = NULL;
ptpcoll_module->super.hier_scather_offset = 0;
/* no header support in ptp */
ptpcoll_module->super.header_size = 0;
/* No network context */
ptpcoll_module->super.network_context = NULL;
/* set the upper limit on the tag */
i = 2;
ptpcoll_module->tag_mask = 1;
while ( i <= (uint64_t) mca_pml.pml_max_tag && i > 0) {
i <<= 1;
}
ptpcoll_module->ml_mem.ml_buf_desc = NULL;
ptpcoll_module->tag_mask = i - 1;
}
static void
mca_bcol_ptpcoll_module_destruct(mca_bcol_ptpcoll_module_t *ptpcoll_module)
{
int i;
mca_bcol_ptpcoll_local_mlmem_desc_t *ml_mem = &ptpcoll_module->ml_mem;
if (NULL != ml_mem->ml_buf_desc) {
/* Release the memory structs that were cache ML memory data */
uint32_t i, j, ci;
for (i = 0; i < ml_mem->num_banks; i++) {
for (j = 0; j < ml_mem->num_buffers_per_bank; j++) {
ci = i * ml_mem->num_buffers_per_bank + j;
if (NULL != ml_mem->ml_buf_desc[ci].requests) {
free(ml_mem->ml_buf_desc[ci].requests);
}
}
}
/* release the buffer descriptor */
free(ml_mem->ml_buf_desc);
}
if (NULL != ptpcoll_module->allgather_offsets) {
free_allreduce_offsets_array(ptpcoll_module);
}
if (NULL != ptpcoll_module->narray_node) {
for (i = 0; i < ptpcoll_module->group_size; i++) {
if (NULL != ptpcoll_module->narray_node[i].children_ranks) {
free(ptpcoll_module->narray_node[i].children_ranks);
}
}
free(ptpcoll_module->narray_node);
}
OBJ_DESTRUCT(&ptpcoll_module->collreqs_free);
if( NULL != ptpcoll_module->super.list_n_connected ){
free(ptpcoll_module->super.list_n_connected);
ptpcoll_module->super.list_n_connected = NULL;
}
}
OBJ_CLASS_INSTANCE(mca_bcol_ptpcoll_module_t,
mca_bcol_base_module_t,
mca_bcol_ptpcoll_module_construct,
mca_bcol_ptpcoll_module_destruct);
static int init_ml_buf_desc(mca_bcol_ptpcoll_ml_buffer_desc_t **desc, void *base_addr, uint32_t num_banks,
uint32_t num_buffers_per_bank, uint32_t size_buffer, uint32_t header_size, int group_size, int pow_k)
{
uint32_t i, j, ci;
mca_bcol_ptpcoll_ml_buffer_desc_t *tmp_desc = NULL;
int k_nomial_radix = mca_bcol_ptpcoll_component.k_nomial_radix;
int pow_k_val = (0 == pow_k) ? 1 : pow_k;
int num_to_alloc =
((k_nomial_radix - 1) * pow_k_val * 2 + 1 > mca_bcol_ptpcoll_component.narray_radix) ?
(k_nomial_radix - 1) * pow_k_val * 2 + 1 :
mca_bcol_ptpcoll_component.narray_radix * 2;
*desc = (mca_bcol_ptpcoll_ml_buffer_desc_t *)calloc(num_banks * num_buffers_per_bank,
sizeof(mca_bcol_ptpcoll_ml_buffer_desc_t));
if (NULL == *desc) {
PTPCOLL_ERROR(("Failed to allocate memory"));
return OMPI_ERROR;
}
tmp_desc = *desc;
for (i = 0; i < num_banks; i++) {
for (j = 0; j < num_buffers_per_bank; j++) {
ci = i * num_buffers_per_bank + j;
tmp_desc[ci].bank_index = i;
tmp_desc[ci].buffer_index = j;
/* *2 is for gather session +1 for extra peer */
tmp_desc[ci].requests = (ompi_request_t **)
calloc(num_to_alloc, sizeof(ompi_request_t *));
if (NULL == tmp_desc[ci].requests) {
PTPCOLL_ERROR(("Failed to allocate memory for requests"));
return OMPI_ERROR;
}
/*
* ptpcoll don't have any header, but other bcols may to have. So
* we need to take it in account.
*/
tmp_desc[ci].data_addr = (void *)
((unsigned char*)base_addr + ci * size_buffer + header_size);
PTPCOLL_VERBOSE(10, ("ml memory cache setup %d %d - %p", i, j, tmp_desc[ci].data_addr));
/* init reduce implementation flags */
tmp_desc[ci].reduce_init_called = false;
tmp_desc[ci].reduction_status = 0;
}
}
return OMPI_SUCCESS;
}
static void mca_bcol_ptpcoll_set_small_msg_thresholds(struct mca_bcol_base_module_t *super)
{
mca_bcol_ptpcoll_module_t *ptpcoll_module =
(mca_bcol_ptpcoll_module_t *) super;
mca_bcol_ptpcoll_component_t *cm = &mca_bcol_ptpcoll_component;
/* Subtract out the maximum header size when calculating the thresholds. This
* will account for the headers used by the basesmuma component. If we do not
* take these headers into account we may overrun our buffer. */
/* Set the Allgather threshold equals to a ML buff size */
super->small_message_thresholds[BCOL_ALLGATHER] =
(ptpcoll_module->ml_mem.size_buffer - BCOL_HEADER_MAX) /
ompi_comm_size(ptpcoll_module->super.sbgp_partner_module->group_comm);
/* Set the Bcast threshold, all Bcast algths have the same threshold */
super->small_message_thresholds[BCOL_BCAST] =
(ptpcoll_module->ml_mem.size_buffer - BCOL_HEADER_MAX);
/* Set the Alltoall threshold, the Ring algth sets some limitation */
super->small_message_thresholds[BCOL_ALLTOALL] =
(ptpcoll_module->ml_mem.size_buffer - BCOL_HEADER_MAX) / 2;
/* Set the Allreduce threshold, NARRAY algth sets some limitation */
super->small_message_thresholds[BCOL_ALLREDUCE] =
(ptpcoll_module->ml_mem.size_buffer - BCOL_HEADER_MAX) / ptpcoll_module->k_nomial_radix;
/* Set the Reduce threshold, NARRAY algth sets some limitation */
super->small_message_thresholds[BCOL_REDUCE] =
(ptpcoll_module->ml_mem.size_buffer - BCOL_HEADER_MAX) / cm->narray_radix;
}
/*
* Cache information about ML memory
*/
static int mca_bcol_ptpcoll_cache_ml_memory_info(struct mca_coll_ml_module_t *ml_module,
struct mca_bcol_base_module_t *bcol,
void *reg_data)
{
mca_bcol_ptpcoll_module_t *ptpcoll_module = (mca_bcol_ptpcoll_module_t *) bcol;
mca_bcol_ptpcoll_local_mlmem_desc_t *ml_mem = &ptpcoll_module->ml_mem;
struct ml_memory_block_desc_t *desc = ml_module->payload_block;
int group_size = ptpcoll_module->super.sbgp_partner_module->group_size;
PTPCOLL_VERBOSE(10, ("mca_bcol_ptpcoll_init_buffer_memory was called"));
/* cache ml mem desc tunings localy */
ml_mem->num_banks = desc->num_banks;
ml_mem->num_buffers_per_bank = desc->num_buffers_per_bank;
ml_mem->size_buffer = desc->size_buffer;
PTPCOLL_VERBOSE(10, ("ML buffer configuration num banks %d num_per_bank %d size %d base addr %p",
desc->num_banks, desc->num_buffers_per_bank, desc->size_buffer, desc->block->base_addr));
/* pointer to ml level descriptor */
ml_mem->ml_mem_desc = desc;
/* Set first bank index for release */
ml_mem->bank_index_for_release = 0;
if (OMPI_SUCCESS != init_ml_buf_desc(&ml_mem->ml_buf_desc,
desc->block->base_addr,
ml_mem->num_banks,
ml_mem->num_buffers_per_bank,
ml_mem->size_buffer,
ml_module->data_offset,
group_size,
ptpcoll_module->pow_k)) {
PTPCOLL_VERBOSE(10, ("Failed to allocate rdma memory descriptor\n"));
return OMPI_ERROR;
}
PTPCOLL_VERBOSE(10, ("ml_module = %p, ptpcoll_module = %p, ml_mem_desc = %p.\n",
ml_module, ptpcoll_module, ml_mem->ml_mem_desc));
return OMPI_SUCCESS;
}
/*
* Load ptpcoll bcol functions
*/
static void load_func(mca_bcol_ptpcoll_module_t *ptpcoll_module)
{
int fnc;
/* reset everything to NULL */
for (fnc = 0; fnc < BCOL_NUM_OF_FUNCTIONS; fnc++) {
/*ptpcoll_module->super.bcol_function_table[fnc] = NULL;*/
ptpcoll_module->super.bcol_function_table[fnc] = NULL;
ptpcoll_module->super.bcol_function_init_table[fnc] = NULL;
}
ptpcoll_module->super.bcol_function_init_table[BCOL_BARRIER] = bcol_ptpcoll_barrier_init;
ptpcoll_module->super.bcol_function_init_table[BCOL_BCAST] = bcol_ptpcoll_bcast_init;
ptpcoll_module->super.bcol_function_init_table[BCOL_ALLREDUCE] = bcol_ptpcoll_allreduce_init;
ptpcoll_module->super.bcol_function_init_table[BCOL_ALLGATHER] = bcol_ptpcoll_allgather_init;
ptpcoll_module->super.bcol_function_table[BCOL_BCAST] = bcol_ptpcoll_bcast_k_nomial_anyroot;
ptpcoll_module->super.bcol_function_init_table[BCOL_ALLTOALL] = NULL;
ptpcoll_module->super.bcol_function_init_table[BCOL_SYNC] = mca_bcol_ptpcoll_memsync_init;
ptpcoll_module->super.bcol_function_init_table[BCOL_REDUCE] = bcol_ptpcoll_reduce_init;
/* ML memory cacher */
ptpcoll_module->super.bcol_memory_init = mca_bcol_ptpcoll_cache_ml_memory_info;
/* Set thresholds */
ptpcoll_module->super.set_small_msg_thresholds = mca_bcol_ptpcoll_set_small_msg_thresholds;
/* setup recursive k-ing tree */
ptpcoll_module->super.k_nomial_tree = mca_bcol_ptpcoll_setup_knomial_tree;
}
int mca_bcol_ptpcoll_setup_knomial_tree(mca_bcol_base_module_t *super)
{
mca_bcol_ptpcoll_module_t *p2p_module = (mca_bcol_ptpcoll_module_t *) super;
int rc = 0;
rc = netpatterns_setup_recursive_knomial_allgather_tree_node(
p2p_module->super.sbgp_partner_module->group_size,
p2p_module->super.sbgp_partner_module->my_index,
mca_bcol_ptpcoll_component.k_nomial_radix,
super->list_n_connected,
&p2p_module->knomial_allgather_tree);
return rc;
}
/* The function used to calculate size */
static int calc_full_tree_size(int radix, int group_size, int *num_leafs)
{
int level_cnt = 1;
int total_cnt = 0;
while( total_cnt < group_size ) {
total_cnt += level_cnt;
level_cnt *= radix;
}
if (total_cnt > group_size) {
*num_leafs = level_cnt / radix;
return total_cnt - level_cnt / radix;
} else {
*num_leafs = level_cnt;
return group_size;
}
}
/* Setup N-array scatter Knomial-gather static information */
static int load_narray_knomial_tree (mca_bcol_ptpcoll_module_t *ptpcoll_module)
{
int rc, i, peer;
mca_bcol_ptpcoll_component_t *cm = &mca_bcol_ptpcoll_component;
ptpcoll_module->full_narray_tree_size = calc_full_tree_size(
cm->narray_knomial_radix,
ptpcoll_module->group_size,
&ptpcoll_module->full_narray_tree_num_leafs);
ptpcoll_module->narray_knomial_proxy_extra_index = (int *)
malloc(sizeof(int) * (cm->narray_knomial_radix));
if (NULL == ptpcoll_module->narray_knomial_proxy_extra_index) {
PTPCOLL_ERROR(("Failed to allocate memory"));
goto Error;
}
ptpcoll_module->narray_knomial_node = calloc(
ptpcoll_module->full_narray_tree_size,
sizeof(netpatterns_narray_knomial_tree_node_t));
if(NULL == ptpcoll_module->narray_knomial_node) {
goto Error;
}
PTPCOLL_VERBOSE(10 ,("My type is proxy, full tree size = %d [%d]",
ptpcoll_module->full_narray_tree_size,
cm->narray_knomial_radix
));
if (ptpcoll_module->super.sbgp_partner_module->my_index <
ptpcoll_module->full_narray_tree_size) {
if (ptpcoll_module->super.sbgp_partner_module->my_index <
ptpcoll_module->group_size - ptpcoll_module->full_narray_tree_size) {
ptpcoll_module->narray_type = PTPCOLL_PROXY;
for (i = 0; i < cm->narray_knomial_radix; i++) {
peer =
ptpcoll_module->super.sbgp_partner_module->my_index *
cm->narray_knomial_radix + i +
ptpcoll_module->full_narray_tree_size;
if (peer >= ptpcoll_module->group_size) {
break;
}
ptpcoll_module->narray_knomial_proxy_extra_index[i] = peer;
}
ptpcoll_module->narray_knomial_proxy_num = i;
} else {
ptpcoll_module->narray_type = PTPCOLL_IN_GROUP;;
}
/* Setting node info */
for(i = 0; i < ptpcoll_module->full_narray_tree_size; i++) {
rc = netpatterns_setup_narray_knomial_tree(
cm->narray_knomial_radix,
i,
ptpcoll_module->full_narray_tree_size,
&ptpcoll_module->narray_knomial_node[i]);
if(OMPI_SUCCESS != rc) {
goto Error;
}
}
} else {
ptpcoll_module->narray_type = PTPCOLL_EXTRA;
ptpcoll_module->narray_knomial_proxy_extra_index[0] =
(ptpcoll_module->super.sbgp_partner_module->my_index -
ptpcoll_module->full_narray_tree_size) /
cm->narray_knomial_radix;
}
return OMPI_SUCCESS;
Error:
if (NULL != ptpcoll_module->narray_knomial_node) {
free(ptpcoll_module->narray_knomial_node);
}
if (NULL != ptpcoll_module->narray_knomial_proxy_extra_index) {
free(ptpcoll_module->narray_knomial_proxy_extra_index);
}
return OMPI_ERROR;
}
/* Setup N-array static information */
static int load_narray_tree(mca_bcol_ptpcoll_module_t *ptpcoll_module)
{
int rc, i;
mca_bcol_ptpcoll_component_t *cm = &mca_bcol_ptpcoll_component;
ptpcoll_module->narray_node = calloc(ptpcoll_module->group_size,
sizeof(netpatterns_tree_node_t));
if(NULL == ptpcoll_module->narray_node ) {
goto Error;
}
for(i = 0; i < ptpcoll_module->group_size; i++) {
rc = netpatterns_setup_narray_tree(
cm->narray_radix,
i,
ptpcoll_module->group_size,
&ptpcoll_module->narray_node[i]);
if(OMPI_SUCCESS != rc) {
goto Error;
}
}
return OMPI_SUCCESS;
Error:
if (NULL != ptpcoll_module->narray_node) {
free(ptpcoll_module->narray_node);
}
return OMPI_ERROR;
}
static int load_knomial_info(mca_bcol_ptpcoll_module_t *ptpcoll_module)
{
int i;
mca_bcol_ptpcoll_component_t *cm = &mca_bcol_ptpcoll_component;
ptpcoll_module->k_nomial_radix =
cm->k_nomial_radix > ptpcoll_module->group_size ?
ptpcoll_module->group_size :
cm->k_nomial_radix;
ptpcoll_module->pow_k = pow_k_calc(ptpcoll_module->k_nomial_radix,
ptpcoll_module->group_size,
&ptpcoll_module->pow_knum);
ptpcoll_module->kn_proxy_extra_index = (int *)
malloc(sizeof(int) * (ptpcoll_module->k_nomial_radix - 1));
if (NULL == ptpcoll_module->kn_proxy_extra_index) {
PTPCOLL_ERROR(("Failed to allocate memory"));
goto Error;
}
/* Setting peer type for K-nomial algorithm*/
if (ptpcoll_module->super.sbgp_partner_module->my_index < ptpcoll_module->pow_knum ) {
if (ptpcoll_module->super.sbgp_partner_module->my_index <
ptpcoll_module->group_size - ptpcoll_module->pow_knum) {
for (i = 0;
i < (ptpcoll_module->k_nomial_radix - 1) &&
ptpcoll_module->super.sbgp_partner_module->my_index *
(ptpcoll_module->k_nomial_radix - 1) +
i + ptpcoll_module->pow_knum < ptpcoll_module->group_size
; i++) {
ptpcoll_module->pow_ktype = PTPCOLL_KN_PROXY;
ptpcoll_module->kn_proxy_extra_index[i] =
ptpcoll_module->super.sbgp_partner_module->my_index *
(ptpcoll_module->k_nomial_radix - 1) +
i + ptpcoll_module->pow_knum;
PTPCOLL_VERBOSE(10 ,("My type is proxy, pow_knum = %d [%d] my extra %d",
ptpcoll_module->pow_knum,
ptpcoll_module->pow_k,
ptpcoll_module->kn_proxy_extra_index[i]));
}
ptpcoll_module->kn_proxy_extra_num = i;
} else {
PTPCOLL_VERBOSE(10 ,("My type is in group, pow_knum = %d [%d]", ptpcoll_module->pow_knum,
ptpcoll_module->pow_k));
ptpcoll_module->pow_ktype = PTPCOLL_KN_IN_GROUP;
}
} else {
ptpcoll_module->pow_ktype = PTPCOLL_KN_EXTRA;
ptpcoll_module->kn_proxy_extra_index[0] = (ptpcoll_module->super.sbgp_partner_module->my_index -
ptpcoll_module->pow_knum) / (ptpcoll_module->k_nomial_radix - 1);
PTPCOLL_VERBOSE(10 ,("My type is extra , pow_knum = %d [%d] my proxy %d",
ptpcoll_module->pow_knum,
ptpcoll_module->pow_k,
ptpcoll_module->kn_proxy_extra_index[0]));
}
return OMPI_SUCCESS;
Error:
if (NULL == ptpcoll_module->kn_proxy_extra_index) {
free(ptpcoll_module->kn_proxy_extra_index);
}
return OMPI_ERROR;
}
static int load_binomial_info(mca_bcol_ptpcoll_module_t *ptpcoll_module)
{
ptpcoll_module->pow_2 = pow_k_calc(2,
ptpcoll_module->group_size,
&ptpcoll_module->pow_2num);
assert(ptpcoll_module->pow_2num == 1 << ptpcoll_module->pow_2);
assert(ptpcoll_module->pow_2num <= ptpcoll_module->group_size);
/* Setting peer type for binary algorithm*/
if (ptpcoll_module->super.sbgp_partner_module->my_index < ptpcoll_module->pow_2num ) {
if (ptpcoll_module->super.sbgp_partner_module->my_index <
ptpcoll_module->group_size - ptpcoll_module->pow_2num) {
PTPCOLL_VERBOSE(10 ,("My type is proxy, pow_2num = %d [%d]", ptpcoll_module->pow_2num,
ptpcoll_module->pow_2));
ptpcoll_module->pow_2type = PTPCOLL_PROXY;
ptpcoll_module->proxy_extra_index = ptpcoll_module->super.sbgp_partner_module->my_index +
ptpcoll_module->pow_2num;
} else {
PTPCOLL_VERBOSE(10 ,("My type is in group, pow_2num = %d [%d]", ptpcoll_module->pow_2num,
ptpcoll_module->pow_2));
ptpcoll_module->pow_2type = PTPCOLL_IN_GROUP;
}
} else {
PTPCOLL_VERBOSE(10 ,("My type is extra , pow_2num = %d [%d]", ptpcoll_module->pow_2num,
ptpcoll_module->pow_2));
ptpcoll_module->pow_2type = PTPCOLL_EXTRA;
ptpcoll_module->proxy_extra_index = ptpcoll_module->super.sbgp_partner_module->my_index -
ptpcoll_module->pow_2num;
}
return OMPI_SUCCESS;
}
static int load_recursive_knomial_info(mca_bcol_ptpcoll_module_t *ptpcoll_module)
{
int rc = OMPI_SUCCESS;
rc = netpatterns_setup_recursive_knomial_tree_node(
ptpcoll_module->group_size,
ptpcoll_module->super.sbgp_partner_module->my_index,
mca_bcol_ptpcoll_component.k_nomial_radix,
&ptpcoll_module->knomial_exchange_tree);
return rc;
}
static void bcol_ptpcoll_collreq_init(ompi_free_list_item_t *item, void* ctx)
{
mca_bcol_ptpcoll_module_t *ptpcoll_module= (mca_bcol_ptpcoll_module_t *) ctx;
mca_bcol_ptpcoll_collreq_t *collreq = (mca_bcol_ptpcoll_collreq_t *) item;
switch(mca_bcol_ptpcoll_component.barrier_alg) {
case 1:
collreq->requests = (ompi_request_t **)
calloc(2, sizeof(ompi_request_t *));
break;
case 2:
collreq->requests = (ompi_request_t **)
calloc(2 * ptpcoll_module->k_nomial_radix, sizeof(ompi_request_t *));
break;
}
}
/* query to see if the module is available for use on the given
* communicator, and if so, what it's priority is. This is where
* the backing shared-memory file is created.
*/
mca_bcol_base_module_t **mca_bcol_ptpcoll_comm_query(mca_sbgp_base_module_t *sbgp,
int *num_modules)
{
int rc;
/* local variables */
struct ompi_communicator_t *comm = sbgp->group_comm;
mca_bcol_ptpcoll_module_t *ptpcoll_module = NULL;
mca_bcol_base_module_t **ptpcoll_modules = NULL;
int iovec_size;
/* initialize local variables */
*num_modules = 0;
/*
* This is activated only for intra-communicators
*/
if (OMPI_COMM_IS_INTER(comm) ) {
return NULL;
}
/* allocate and initialize an sm-v2 module */
ptpcoll_modules = (mca_bcol_base_module_t **) malloc(sizeof(mca_bcol_base_module_t *));
if (NULL == ptpcoll_modules) {
return NULL;
}
ptpcoll_module = OBJ_NEW(mca_bcol_ptpcoll_module_t);
if (NULL == ptpcoll_module) {
return NULL;
}
/* On this stage we support only one single module */
ptpcoll_modules[*num_modules] = &(ptpcoll_module->super);
(*num_modules)++;
/* set the subgroup */
ptpcoll_module->super.sbgp_partner_module = sbgp;
/* caching some useful information */
ptpcoll_module->group_size =
ptpcoll_module->super.sbgp_partner_module->group_size;
rc = load_binomial_info(ptpcoll_module);
if (OMPI_SUCCESS != rc) {
PTPCOLL_VERBOSE(10, ("Failed to load knomial info"));
goto CLEANUP;
}
rc = load_knomial_info(ptpcoll_module);
if (OMPI_SUCCESS != rc) {
PTPCOLL_VERBOSE(10, ("Failed to load knomial info"));
goto CLEANUP;
}
rc = load_narray_tree(ptpcoll_module);
if (OMPI_SUCCESS != rc) {
PTPCOLL_VERBOSE(10, ("Failed to load narray tree"));
goto CLEANUP;
}
rc = load_narray_knomial_tree(ptpcoll_module);
if (OMPI_SUCCESS != rc) {
PTPCOLL_VERBOSE(10, ("Failed to load narray-knomila tree"));
goto CLEANUP;
}
rc = load_recursive_knomial_info(ptpcoll_module);
if (OMPI_SUCCESS != rc) {
PTPCOLL_VERBOSE(10, ("Failed to load recursive knomial tree"));
goto CLEANUP;
}
/* creating collfrag free list */
OBJ_CONSTRUCT(&ptpcoll_module->collreqs_free, ompi_free_list_t);
rc = ompi_free_list_init_ex_new(&ptpcoll_module->collreqs_free,
sizeof(mca_bcol_ptpcoll_collreq_t),
BCOL_PTP_CACHE_LINE_SIZE,
OBJ_CLASS(mca_bcol_ptpcoll_collreq_t),
0, BCOL_PTP_CACHE_LINE_SIZE,
256 /* free_list_num */,
-1 /* free_list_max, -1 = infinite */,
32 /* free_list_inc */,
NULL,
bcol_ptpcoll_collreq_init,
ptpcoll_module);
if (OMPI_SUCCESS != rc) {
goto CLEANUP;
}
load_func(ptpcoll_module);
rc = alloc_allreduce_offsets_array(ptpcoll_module);
if (OMPI_SUCCESS != rc) {
goto CLEANUP;
}
/* Allocating iovec for PTP alltoall */
iovec_size = ptpcoll_module->group_size / 2 + ptpcoll_module->group_size % 2;
ptpcoll_module->alltoall_iovec = (struct iovec *) malloc(sizeof(struct iovec)
* iovec_size);
ptpcoll_module->log_group_size = lognum(ptpcoll_module->group_size);
rc = mca_bcol_base_bcol_fns_table_init(&(ptpcoll_module->super));
if (OMPI_SUCCESS != rc) {
goto CLEANUP;
}
/* Zero copy is supported */
ptpcoll_module->super.supported_mode = MCA_BCOL_BASE_ZERO_COPY;
/* return */
return ptpcoll_modules;
CLEANUP:
OBJ_RELEASE(ptpcoll_module);
return NULL;
}
Показать git blame Копировать постоянную ссылку