ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность

* add ability to try a couple of different collectives for fence

Просмотр исходного кода 
synchronization to see which gives the best performance

This commit was SVN r9314.
Этот коммит содержится в:
Brian Barrett 2006-03-16 18:40:42 +00:00
родитель 8a9e76dfa3
Коммит 234adb292b
3 изменённых файлов: 98 добавлений и 14 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

8
ompi/mca/osc/pt2pt/osc_pt2pt.h
Просмотреть файл

@ -109,6 +109,11 @@ struct ompi_osc_pt2pt_module_t {
/* an array of <sizeof(p2p_comm)> ints, each containing the value
1. */
int *p2p_fence_coll_counts;
/* an array of <sizeof(p2p_comm)> shorts, for use in experimenting
with different synchronization costs */
short *p2p_fence_coll_results;
enum { OSC_SYNC_REDUCE_SCATTER, OSC_SYNC_ALLREDUCE, OSC_SYNC_ALLTOALL } p2p_fence_sync_type;
/* ********************* PWSC data ************************ */
@ -163,9 +168,6 @@ extern "C" {
/*
* Component functions
*/
int ompi_osc_pt2pt_component_open(void);
int ompi_osc_pt2pt_component_close(void);
int ompi_osc_pt2pt_component_init(bool enable_progress_threads,
bool enable_mpi_threads);

49
ompi/mca/osc/pt2pt/osc_pt2pt_component.c
Просмотреть файл

@ -34,6 +34,8 @@
#include "ompi/mca/bml/base/base.h"
#include "ompi/datatype/dt_arch.h"
static int ompi_osc_pt2pt_component_open(void);
ompi_osc_pt2pt_component_t mca_osc_pt2pt_component = {
{ /* ompi_osc_base_component_t */
{ /* ompi_base_component_t */
@ -42,7 +44,7 @@ ompi_osc_pt2pt_component_t mca_osc_pt2pt_component = {
1,
0,
0,
NULL,
ompi_osc_pt2pt_component_open,
NULL
},
{ /* mca_base_component_data */
@ -111,6 +113,20 @@ want_locks(ompi_info_t *info)
return !no_locks;
}
static int fence_sync_index;
static int
ompi_osc_pt2pt_component_open(void)
{
fence_sync_index =
mca_base_param_reg_string(&mca_osc_pt2pt_component.super.osc_version,
"fence_sync_method",
"How to synchronize fence: reduce_scatter, allreduce, alltoall",
false, false, "reduce_scatter", NULL);
return OMPI_SUCCESS;
}
int
ompi_osc_pt2pt_component_init(bool enable_progress_threads,
@ -195,6 +211,8 @@ ompi_osc_pt2pt_component_select(ompi_win_t *win,
{
ompi_osc_pt2pt_module_t *module;
int ret, i;
char *sync_string;
/* create module structure */
module = malloc(sizeof(ompi_osc_pt2pt_module_t));
if (NULL == module) return OMPI_ERROR;
@ -273,6 +291,35 @@ ompi_osc_pt2pt_component_select(ompi_win_t *win,
module->p2p_fence_coll_counts[i] = 1;
}
module->p2p_fence_coll_results = malloc(sizeof(int) *
ompi_comm_size(module->p2p_comm));
if (NULL == module->p2p_fence_coll_counts) {
free(module->p2p_fence_coll_counts);
free(module->p2p_copy_num_pending_sendreqs);
OBJ_DESTRUCT(&module->p2p_copy_pending_sendreqs);
OBJ_DESTRUCT(&module->p2p_long_msgs);
free(module->p2p_num_pending_sendreqs);
OBJ_DESTRUCT(&module->p2p_pending_sendreqs);
ompi_comm_free(&comm);
OBJ_DESTRUCT(&(module->p2p_acc_lock));
OBJ_DESTRUCT(&(module->p2p_lock));
free(module);
return ret;
}
/* figure out what sync method to use */
mca_base_param_lookup_string(fence_sync_index, &sync_string);
if (0 == strcmp(sync_string, "reduce_scatter")) {
module->p2p_fence_sync_type = OSC_SYNC_REDUCE_SCATTER;
} else if (0 == strcmp(sync_string, "allreduce")) {
module->p2p_fence_sync_type = OSC_SYNC_ALLREDUCE;
} else if (0 == strcmp(sync_string, "alltoall")) {
module->p2p_fence_sync_type = OSC_SYNC_ALLTOALL;
} else {
opal_output(0, "invalid value for fence_sync_method parameter: %s\n", sync_string);
return OMPI_ERROR;
}
/* pwsc data */
module->p2p_pw_group = NULL;
module->p2p_sc_group = NULL;

55
ompi/mca/osc/pt2pt/osc_pt2pt_sync.c
Просмотреть файл

@ -112,16 +112,51 @@ ompi_osc_pt2pt_module_fence(int assert, ompi_win_t *win)
ompi_osc_pt2pt_flip_sendreqs(P2P_MODULE(win));
/* find out how much data everyone is going to send us. Need
to have the lock during this period so that we have a sane
view of the number of sendreqs */
ret = P2P_MODULE(win)->p2p_comm->
c_coll.coll_reduce_scatter(P2P_MODULE(win)->p2p_copy_num_pending_sendreqs,
&incoming_reqs,
P2P_MODULE(win)->p2p_fence_coll_counts,
MPI_SHORT,
MPI_SUM,
P2P_MODULE(win)->p2p_comm);
switch (P2P_MODULE(win)->p2p_fence_sync_type) {
/* find out how much data everyone is going to send us. Need
to have the lock during this period so that we have a sane
view of the number of sendreqs */
case OSC_SYNC_REDUCE_SCATTER:
ret = P2P_MODULE(win)->p2p_comm->
c_coll.coll_reduce_scatter(P2P_MODULE(win)->p2p_copy_num_pending_sendreqs,
&incoming_reqs,
P2P_MODULE(win)->p2p_fence_coll_counts,
MPI_SHORT,
MPI_SUM,
P2P_MODULE(win)->p2p_comm);
break;
case OSC_SYNC_ALLREDUCE:
ret = P2P_MODULE(win)->p2p_comm->
c_coll.coll_allreduce(P2P_MODULE(win)->p2p_copy_num_pending_sendreqs,
P2P_MODULE(win)->p2p_fence_coll_results,
ompi_comm_size(P2P_MODULE(win)->p2p_comm),
MPI_SHORT,
MPI_SUM,
P2P_MODULE(win)->p2p_comm);
incoming_reqs = P2P_MODULE(win)->
p2p_fence_coll_results[P2P_MODULE(win)->p2p_comm->c_my_rank];
break;
case OSC_SYNC_ALLTOALL:
ret = P2P_MODULE(win)->p2p_comm->
c_coll.coll_alltoall(P2P_MODULE(win)->p2p_copy_num_pending_sendreqs,
1,
MPI_SHORT,
P2P_MODULE(win)->p2p_fence_coll_results,
1,
MPI_SHORT,
P2P_MODULE(win)->p2p_comm);
incoming_reqs = 0;
for (i = 0 ; i < ompi_comm_size(P2P_MODULE(win)->p2p_comm) ; ++i) {
incoming_reqs += P2P_MODULE(win)->p2p_fence_coll_results[i];
}
break;
default:
assert(0 == 1);
}
if (OMPI_SUCCESS != ret) {
/* put the stupid data back for the user. This is not
cheap, but the user lost his data if we don't. */