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

Merge pull request #925 from hjelmn/osc_sm_fix

Просмотр исходного кода 
osc/sm: fix pscw synchronization
Этот коммит содержится в:
Nathan Hjelm 2015-09-22 19:14:08 -06:00
родитель 5553dba0c4 ee5810813b
Коммит b80cd56326
6 изменённых файлов: 305 добавлений и 151 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

38
ompi/mca/osc/pt2pt/osc_pt2pt_active_target.c
Просмотреть файл

@ -114,12 +114,14 @@ static ompi_osc_pt2pt_peer_t **ompi_osc_pt2pt_get_peers (ompi_osc_pt2pt_module_t
static void ompi_osc_pt2pt_release_peers (ompi_osc_pt2pt_peer_t **peers, int npeers) static void ompi_osc_pt2pt_release_peers (ompi_osc_pt2pt_peer_t **peers, int npeers)
{ {
if (peers) {
for (int i = 0 ; i < npeers ; ++i) { for (int i = 0 ; i < npeers ; ++i) {
OBJ_RELEASE(peers[i]); OBJ_RELEASE(peers[i]);
} }
free (peers); free (peers);
} }
}
int ompi_osc_pt2pt_fence(int assert, ompi_win_t *win) int ompi_osc_pt2pt_fence(int assert, ompi_win_t *win)
{ {
@ -228,20 +230,24 @@ int ompi_osc_pt2pt_start (ompi_group_t *group, int assert, ompi_win_t *win)
"ompi_osc_pt2pt_start entering with group size %d...", "ompi_osc_pt2pt_start entering with group size %d...",
sync->num_peers)); sync->num_peers));
if (0 == ompi_group_size (group)) {
/* nothing more to do. this is an empty start epoch */
OPAL_THREAD_UNLOCK(&module->lock);
return OMPI_SUCCESS;
}
opal_atomic_wmb ();
sync->type = OMPI_OSC_PT2PT_SYNC_TYPE_PSCW; sync->type = OMPI_OSC_PT2PT_SYNC_TYPE_PSCW;
/* prevent us from entering a passive-target, fence, or another pscw access epoch until /* prevent us from entering a passive-target, fence, or another pscw access epoch until
* the matching complete is called */ * the matching complete is called */
sync->epoch_active = true; sync->epoch_active = true;
/* save the group */
OBJ_RETAIN(group);
if (0 == ompi_group_size (group)) {
/* nothing more to do. this is an empty start epoch */
sync->eager_send_active = true;
OPAL_THREAD_UNLOCK(&module->lock);
return OMPI_SUCCESS;
}
opal_atomic_wmb ();
/* translate the group ranks into the communicator */ /* translate the group ranks into the communicator */
sync->peer_list.peers = ompi_osc_pt2pt_get_peers (module, group); sync->peer_list.peers = ompi_osc_pt2pt_get_peers (module, group);
if (NULL == sync->peer_list.peers) { if (NULL == sync->peer_list.peers) {
@ -249,10 +255,6 @@ int ompi_osc_pt2pt_start (ompi_group_t *group, int assert, ompi_win_t *win)
return OMPI_ERR_OUT_OF_RESOURCE; return OMPI_ERR_OUT_OF_RESOURCE;
} }
/* save the group */
OBJ_RETAIN(group);
ompi_group_increment_proc_count(group);
if (!(assert & MPI_MODE_NOCHECK)) { if (!(assert & MPI_MODE_NOCHECK)) {
OPAL_THREAD_LOCK(&sync->lock); OPAL_THREAD_LOCK(&sync->lock);
for (int i = 0 ; i < sync->num_peers ; ++i) { for (int i = 0 ; i < sync->num_peers ; ++i) {
@ -318,12 +320,6 @@ int ompi_osc_pt2pt_complete (ompi_win_t *win)
group_size = sync->num_peers; group_size = sync->num_peers;
peers = sync->peer_list.peers; peers = sync->peer_list.peers;
if (NULL == peers) {
/* empty peer list */
OPAL_THREAD_UNLOCK(&(module->lock));
OBJ_RELEASE(group);
return OMPI_SUCCESS;
}
OPAL_THREAD_UNLOCK(&module->lock); OPAL_THREAD_UNLOCK(&module->lock);
@ -383,8 +379,10 @@ int ompi_osc_pt2pt_complete (ompi_win_t *win)
module->epoch_outgoing_frag_count[rank] = 0; module->epoch_outgoing_frag_count[rank] = 0;
} }
if (peers) {
/* release our reference to peers in this group */ /* release our reference to peers in this group */
ompi_osc_pt2pt_release_peers (peers, group_size); ompi_osc_pt2pt_release_peers (peers, group_size);
}
if (OMPI_SUCCESS != ret) { if (OMPI_SUCCESS != ret) {
return ret; return ret;
@ -403,7 +401,6 @@ int ompi_osc_pt2pt_complete (ompi_win_t *win)
OPAL_THREAD_UNLOCK(&module->lock); OPAL_THREAD_UNLOCK(&module->lock);
/* phase 2 cleanup group */ /* phase 2 cleanup group */
ompi_group_decrement_proc_count(group);
OBJ_RELEASE(group); OBJ_RELEASE(group);
OPAL_OUTPUT_VERBOSE((50, ompi_osc_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((50, ompi_osc_base_framework.framework_output,
@ -439,7 +436,6 @@ int ompi_osc_pt2pt_post (ompi_group_t *group, int assert, ompi_win_t *win)
/* save the group */ /* save the group */
OBJ_RETAIN(group); OBJ_RETAIN(group);
ompi_group_increment_proc_count(group);
module->pw_group = group; module->pw_group = group;
@ -523,7 +519,6 @@ int ompi_osc_pt2pt_wait (ompi_win_t *win)
module->pw_group = NULL; module->pw_group = NULL;
OPAL_THREAD_UNLOCK(&module->lock); OPAL_THREAD_UNLOCK(&module->lock);
ompi_group_decrement_proc_count(group);
OBJ_RELEASE(group); OBJ_RELEASE(group);
OPAL_OUTPUT_VERBOSE((25, ompi_osc_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((25, ompi_osc_base_framework.framework_output,
@ -561,7 +556,6 @@ int ompi_osc_pt2pt_test (ompi_win_t *win, int *flag)
OPAL_THREAD_UNLOCK(&(module->lock)); OPAL_THREAD_UNLOCK(&(module->lock));
ompi_group_decrement_proc_count(group);
OBJ_RELEASE(group); OBJ_RELEASE(group);
return OMPI_SUCCESS; return OMPI_SUCCESS;

2
ompi/mca/osc/pt2pt/osc_pt2pt_sync.h
Просмотреть файл

@ -173,6 +173,8 @@ static inline void ompi_osc_pt2pt_sync_reset (ompi_osc_pt2pt_sync_t *sync)
sync->type = OMPI_OSC_PT2PT_SYNC_TYPE_NONE; sync->type = OMPI_OSC_PT2PT_SYNC_TYPE_NONE;
sync->eager_send_active = 0; sync->eager_send_active = 0;
sync->epoch_active = 0; sync->epoch_active = 0;
sync->peer_list.peers = NULL;
sync->sync.pscw.group = NULL;
} }
#endif /* OMPI_OSC_PT2PT_SYNC_H */ #endif /* OMPI_OSC_PT2PT_SYNC_H */

51
ompi/mca/osc/rdma/osc_rdma_component.c
Просмотреть файл

@ -61,6 +61,8 @@ static int ompi_osc_rdma_component_select (struct ompi_win_t *win, void **base,
static int ompi_osc_rdma_set_info (struct ompi_win_t *win, struct ompi_info_t *info); static int ompi_osc_rdma_set_info (struct ompi_win_t *win, struct ompi_info_t *info);
static int ompi_osc_rdma_get_info (struct ompi_win_t *win, struct ompi_info_t **info_used); static int ompi_osc_rdma_get_info (struct ompi_win_t *win, struct ompi_info_t **info_used);
static int ompi_osc_rdma_query_btls (ompi_communicator_t *comm, struct mca_btl_base_module_t **btl);
static char *ompi_osc_rdma_btl_names; static char *ompi_osc_rdma_btl_names;
ompi_osc_rdma_component_t mca_osc_rdma_component = { ompi_osc_rdma_component_t mca_osc_rdma_component = {
@ -296,10 +298,16 @@ static int ompi_osc_rdma_component_query (struct ompi_win_t *win, void **base, s
struct ompi_communicator_t *comm, struct ompi_info_t *info, struct ompi_communicator_t *comm, struct ompi_info_t *info,
int flavor) int flavor)
{ {
if (MPI_WIN_FLAVOR_SHARED == flavor) { if (MPI_WIN_FLAVOR_SHARED == flavor) {
return -1; return -1;
} }
if (OMPI_SUCCESS != ompi_osc_rdma_query_btls (comm, NULL)) {
return -1;
}
return mca_osc_rdma_component.priority; return mca_osc_rdma_component.priority;
} }
@ -619,32 +627,31 @@ static int allocate_state_shared (ompi_osc_rdma_module_t *module, void **base, s
return ret; return ret;
} }
static int ompi_osc_rdma_find_rdma_endpoints (ompi_osc_rdma_module_t *module) static int ompi_osc_rdma_query_btls (ompi_communicator_t *comm, struct mca_btl_base_module_t **btl)
{ {
struct mca_btl_base_module_t **possible_btls = NULL; struct mca_btl_base_module_t **possible_btls = NULL;
int comm_size = ompi_comm_size (module->comm); int comm_size = ompi_comm_size (comm);
int rc = OMPI_SUCCESS, max_btls = 0; int rc = OMPI_SUCCESS, max_btls = 0;
unsigned int selected_latency = INT_MAX; unsigned int selected_latency = INT_MAX;
mca_btl_base_selected_module_t *selected_btl; struct mca_btl_base_module_t *selected_btl = NULL;
mca_btl_base_selected_module_t *item;
int *btl_counts = NULL; int *btl_counts = NULL;
char **btls_to_use; char **btls_to_use;
void *tmp; void *tmp;
module->selected_btl = NULL;
btls_to_use = opal_argv_split (ompi_osc_rdma_btl_names, ','); btls_to_use = opal_argv_split (ompi_osc_rdma_btl_names, ',');
if (btls_to_use) { if (btls_to_use) {
/* rdma and atomics are only supported with BTLs at the moment */ /* rdma and atomics are only supported with BTLs at the moment */
OPAL_LIST_FOREACH(selected_btl, &mca_btl_base_modules_initialized, mca_btl_base_selected_module_t) { OPAL_LIST_FOREACH(item, &mca_btl_base_modules_initialized, mca_btl_base_selected_module_t) {
for (int i = 0 ; btls_to_use[i] ; ++i) { for (int i = 0 ; btls_to_use[i] ; ++i) {
if (0 != strcmp (btls_to_use[i], selected_btl->btl_module->btl_component->btl_version.mca_component_name)) { if (0 != strcmp (btls_to_use[i], item->btl_module->btl_component->btl_version.mca_component_name)) {
continue; continue;
} }
if ((selected_btl->btl_module->btl_flags & (MCA_BTL_FLAGS_RDMA)) == MCA_BTL_FLAGS_RDMA && if ((item->btl_module->btl_flags & (MCA_BTL_FLAGS_RDMA)) == MCA_BTL_FLAGS_RDMA &&
(selected_btl->btl_module->btl_flags & (MCA_BTL_FLAGS_ATOMIC_FOPS | MCA_BTL_FLAGS_ATOMIC_OPS))) { (item->btl_module->btl_flags & (MCA_BTL_FLAGS_ATOMIC_FOPS | MCA_BTL_FLAGS_ATOMIC_OPS))) {
if (!module->selected_btl || selected_btl->btl_module->btl_latency < module->selected_btl->btl_latency) { if (!selected_btl || item->btl_module->btl_latency < selected_btl->btl_latency) {
module->selected_btl = selected_btl->btl_module; selected_btl = item->btl_module;
} }
} }
} }
@ -653,14 +660,18 @@ static int ompi_osc_rdma_find_rdma_endpoints (ompi_osc_rdma_module_t *module)
opal_argv_free (btls_to_use); opal_argv_free (btls_to_use);
} }
if (NULL != module->selected_btl) { if (btl) {
*btl = selected_btl;
}
if (NULL != selected_btl) {
OPAL_OUTPUT_VERBOSE((11, ompi_osc_base_framework.framework_output, "Selected BTL: %s", OPAL_OUTPUT_VERBOSE((11, ompi_osc_base_framework.framework_output, "Selected BTL: %s",
module->selected_btl->btl_component->btl_version.mca_component_name)); selected_btl->btl_component->btl_version.mca_component_name));
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
for (int i = 0 ; i < comm_size ; ++i) { for (int i = 0 ; i < comm_size ; ++i) {
ompi_proc_t *proc = ompi_comm_peer_lookup (module->comm, i); ompi_proc_t *proc = ompi_comm_peer_lookup (comm, i);
mca_bml_base_endpoint_t *endpoint; mca_bml_base_endpoint_t *endpoint;
int num_btls, prev_max; int num_btls, prev_max;
@ -733,7 +744,7 @@ static int ompi_osc_rdma_find_rdma_endpoints (ompi_osc_rdma_module_t *module)
} }
if (btl_counts[i] == comm_size && possible_btls[i]->btl_latency < selected_latency) { if (btl_counts[i] == comm_size && possible_btls[i]->btl_latency < selected_latency) {
module->selected_btl = possible_btls[i]; selected_btl = possible_btls[i];
selected_latency = possible_btls[i]->btl_latency; selected_latency = possible_btls[i]->btl_latency;
} }
} }
@ -741,13 +752,17 @@ static int ompi_osc_rdma_find_rdma_endpoints (ompi_osc_rdma_module_t *module)
free (possible_btls); free (possible_btls);
free (btl_counts); free (btl_counts);
if (NULL == module->selected_btl) { if (btl) {
*btl = selected_btl;
}
if (NULL == selected_btl) {
/* no btl = no rdma/atomics */ /* no btl = no rdma/atomics */
return OMPI_ERR_NOT_AVAILABLE; return OMPI_ERR_NOT_AVAILABLE;
} }
OPAL_OUTPUT_VERBOSE((11, ompi_osc_base_framework.framework_output, "Selected BTL: %s", OPAL_OUTPUT_VERBOSE((11, ompi_osc_base_framework.framework_output, "Selected BTL: %s",
module->selected_btl->btl_component->btl_version.mca_component_name)); selected_btl->btl_component->btl_version.mca_component_name));
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
@ -1017,7 +1032,7 @@ static int ompi_osc_rdma_component_select (struct ompi_win_t *win, void **base,
} }
/* find rdma capable endpoints */ /* find rdma capable endpoints */
ret = ompi_osc_rdma_find_rdma_endpoints (module); ret = ompi_osc_rdma_query_btls (module->comm, &module->selected_btl);
if (OMPI_SUCCESS != ret) { if (OMPI_SUCCESS != ret) {
OPAL_OUTPUT_VERBOSE((11, ompi_osc_base_framework.framework_output, "Failed finding RDMA endpoints")); OPAL_OUTPUT_VERBOSE((11, ompi_osc_base_framework.framework_output, "Failed finding RDMA endpoints"));
ompi_osc_rdma_free (win); ompi_osc_rdma_free (win);

4
ompi/mca/osc/sm/osc_sm.h
Просмотреть файл

@ -38,7 +38,6 @@ struct ompi_osc_sm_lock_t {
typedef struct ompi_osc_sm_lock_t ompi_osc_sm_lock_t; typedef struct ompi_osc_sm_lock_t ompi_osc_sm_lock_t;
struct ompi_osc_sm_node_state_t { struct ompi_osc_sm_node_state_t {
int32_t post_count;
int32_t complete_count; int32_t complete_count;
ompi_osc_sm_lock_t lock; ompi_osc_sm_lock_t lock;
opal_atomic_lock_t accumulate_lock; opal_atomic_lock_t accumulate_lock;
@ -81,6 +80,9 @@ struct ompi_osc_sm_module_t {
ompi_osc_sm_global_state_t *global_state; ompi_osc_sm_global_state_t *global_state;
ompi_osc_sm_node_state_t *my_node_state; ompi_osc_sm_node_state_t *my_node_state;
ompi_osc_sm_node_state_t *node_states; ompi_osc_sm_node_state_t *node_states;
uint64_t **posts;
opal_mutex_t lock;
}; };
typedef struct ompi_osc_sm_module_t ompi_osc_sm_module_t; typedef struct ompi_osc_sm_module_t ompi_osc_sm_module_t;

207
ompi/mca/osc/sm/osc_sm_active_target.c
Просмотреть файл

@ -1,7 +1,7 @@
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */ /* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/* /*
* Copyright (c) 2012 Sandia National Laboratories. All rights reserved. * Copyright (c) 2012 Sandia National Laboratories. All rights reserved.
* Copyright (c) 2014 Los Alamos National Security, LLC. All rights * Copyright (c) 2014-2015 Los Alamos National Security, LLC. All rights
* reserved. * reserved.
* Copyright (c) 2014 The University of Tennessee and The University * Copyright (c) 2014 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights * of Tennessee Research Foundation. All rights
@ -22,6 +22,74 @@
#include "osc_sm.h" #include "osc_sm.h"
/**
* compare_ranks:
*
* @param[in] ptra Pointer to integer item
* @param[in] ptrb Pointer to integer item
*
* @returns 0 if *ptra == *ptrb
* @returns -1 if *ptra < *ptrb
* @returns 1 otherwise
*
* This function is used to sort the rank list. It can be removed if
* groups are always in order.
*/
static int compare_ranks (const void *ptra, const void *ptrb)
{
int a = *((int *) ptra);
int b = *((int *) ptrb);
if (a < b) {
return -1;
} else if (a > b) {
return 1;
}
return 0;
}
/**
* ompi_osc_pt2pt_get_comm_ranks:
*
* @param[in] module - OSC PT2PT module
* @param[in] sub_group - Group with ranks to translate
*
* @returns an array of translated ranks on success or NULL on failure
*
* Translate the ranks given in {sub_group} into ranks in the
* communicator used to create {module}.
*/
static int *ompi_osc_sm_group_ranks (ompi_group_t *group, ompi_group_t *sub_group)
{
int size = ompi_group_size(sub_group);
int *ranks1, *ranks2;
int ret;
ranks1 = calloc (size, sizeof(int));
ranks2 = calloc (size, sizeof(int));
if (NULL == ranks1 || NULL == ranks2) {
free (ranks1);
free (ranks2);
return NULL;
}
for (int i = 0 ; i < size ; ++i) {
ranks1[i] = i;
}
ret = ompi_group_translate_ranks (sub_group, size, ranks1, group, ranks2);
free (ranks1);
if (OMPI_SUCCESS != ret) {
free (ranks2);
return NULL;
}
qsort (ranks2, size, sizeof (int), compare_ranks);
return ranks2;
}
int int
ompi_osc_sm_fence(int assert, struct ompi_win_t *win) ompi_osc_sm_fence(int assert, struct ompi_win_t *win)
@ -54,7 +122,6 @@ ompi_osc_sm_fence(int assert, struct ompi_win_t *win)
} }
} }
int int
ompi_osc_sm_start(struct ompi_group_t *group, ompi_osc_sm_start(struct ompi_group_t *group,
int assert, int assert,
@ -62,20 +129,43 @@ ompi_osc_sm_start(struct ompi_group_t *group,
{ {
ompi_osc_sm_module_t *module = ompi_osc_sm_module_t *module =
(ompi_osc_sm_module_t*) win->w_osc_module; (ompi_osc_sm_module_t*) win->w_osc_module;
int my_rank = ompi_comm_rank (module->comm);
OBJ_RETAIN(group);
if (!OPAL_ATOMIC_CMPSET(&module->start_group, NULL, group)) {
OBJ_RELEASE(group);
return OMPI_ERR_RMA_SYNC;
}
if (0 == (assert & MPI_MODE_NOCHECK)) { if (0 == (assert & MPI_MODE_NOCHECK)) {
int size; int size;
OBJ_RETAIN(group); int *ranks = ompi_osc_sm_group_ranks (module->comm->c_local_group, group);
module->start_group = group; if (NULL == ranks) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
size = ompi_group_size(module->start_group); size = ompi_group_size(module->start_group);
while (module->my_node_state->post_count != size) { for (int i = 0 ; i < size ; ++i) {
int rank_byte = ranks[i] >> 6;
uint64_t old, rank_bit = 1 << (ranks[i] & 0x3f);
/* wait for rank to post */
while (!(module->posts[my_rank][rank_byte] & rank_bit)) {
opal_progress(); opal_progress();
opal_atomic_mb(); opal_atomic_mb();
} }
} else {
module->start_group = NULL; opal_atomic_rmb ();
do {
old = module->posts[my_rank][rank_byte];
} while (!opal_atomic_cmpset_64 ((int64_t *) module->posts[my_rank] + rank_byte, old, old ^ rank_bit));
}
free (ranks);
} }
opal_atomic_mb(); opal_atomic_mb();
@ -88,30 +178,33 @@ ompi_osc_sm_complete(struct ompi_win_t *win)
{ {
ompi_osc_sm_module_t *module = ompi_osc_sm_module_t *module =
(ompi_osc_sm_module_t*) win->w_osc_module; (ompi_osc_sm_module_t*) win->w_osc_module;
int gsize, csize; ompi_group_t *group;
int gsize;
/* ensure all memory operations have completed */ /* ensure all memory operations have completed */
opal_atomic_mb(); opal_atomic_mb();
if (NULL != module->start_group) { group = module->start_group;
module->my_node_state->post_count = 0; if (NULL == group || !OPAL_ATOMIC_CMPSET(&module->start_group, group, NULL)) {
return OMPI_ERR_RMA_SYNC;
}
opal_atomic_mb(); opal_atomic_mb();
gsize = ompi_group_size(module->start_group); int *ranks = ompi_osc_sm_group_ranks (module->comm->c_local_group, group);
csize = ompi_comm_size(module->comm); if (NULL == ranks) {
for (int i = 0 ; i < gsize ; ++i) { return OMPI_ERR_OUT_OF_RESOURCE;
for (int j = 0 ; j < csize ; ++j) {
if (ompi_group_peer_lookup(module->start_group, i) ==
ompi_comm_peer_lookup(module->comm, j)) {
(void)opal_atomic_add_32(&module->node_states[j].complete_count, 1);
}
}
} }
OBJ_RELEASE(module->start_group); gsize = ompi_group_size(group);
module->start_group = NULL; for (int i = 0 ; i < gsize ; ++i) {
(void) opal_atomic_add_32(&module->node_states[ranks[i]].complete_count, 1);
} }
free (ranks);
OBJ_RELEASE(group);
opal_atomic_mb(); opal_atomic_mb();
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
@ -124,28 +217,44 @@ ompi_osc_sm_post(struct ompi_group_t *group,
{ {
ompi_osc_sm_module_t *module = ompi_osc_sm_module_t *module =
(ompi_osc_sm_module_t*) win->w_osc_module; (ompi_osc_sm_module_t*) win->w_osc_module;
int gsize, csize; int my_rank = ompi_comm_rank (module->comm);
int my_byte = my_rank >> 6;
uint64_t my_bit = 1 << (my_rank & 0x3f);
int gsize;
OPAL_THREAD_LOCK(&module->lock);
if (NULL != module->post_group) {
OPAL_THREAD_UNLOCK(&module->lock);
return OMPI_ERR_RMA_SYNC;
}
module->post_group = group;
OBJ_RETAIN(group);
if (0 == (assert & MPI_MODE_NOCHECK)) { if (0 == (assert & MPI_MODE_NOCHECK)) {
OBJ_RETAIN(group); int *ranks = ompi_osc_sm_group_ranks (module->comm->c_local_group, group);
module->post_group = group; if (NULL == ranks) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
module->my_node_state->complete_count = 0; module->my_node_state->complete_count = 0;
opal_atomic_mb(); opal_atomic_mb();
gsize = ompi_group_size(module->post_group); gsize = ompi_group_size(module->post_group);
csize = ompi_comm_size(module->comm);
for (int i = 0 ; i < gsize ; ++i) { for (int i = 0 ; i < gsize ; ++i) {
for (int j = 0 ; j < csize ; ++j) { (void) opal_atomic_add_64 ((int64_t *) module->posts[ranks[i]] + my_byte, my_bit);
if (ompi_group_peer_lookup(module->post_group, i) ==
ompi_comm_peer_lookup(module->comm, j)) {
(void)opal_atomic_add_32(&module->node_states[j].post_count, 1);
} }
opal_atomic_wmb ();
free (ranks);
opal_progress ();
} }
}
} else { OPAL_THREAD_UNLOCK(&module->lock);
module->post_group = NULL;
}
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
@ -156,18 +265,28 @@ ompi_osc_sm_wait(struct ompi_win_t *win)
{ {
ompi_osc_sm_module_t *module = ompi_osc_sm_module_t *module =
(ompi_osc_sm_module_t*) win->w_osc_module; (ompi_osc_sm_module_t*) win->w_osc_module;
ompi_group_t *group;
if (NULL != module->post_group) { OPAL_THREAD_LOCK(&module->lock);
int size = ompi_group_size(module->post_group);
if (NULL == module->post_group) {
OPAL_THREAD_UNLOCK(&module->lock);
return OMPI_ERR_RMA_SYNC;
}
group = module->post_group;
int size = ompi_group_size (group);
while (module->my_node_state->complete_count != size) { while (module->my_node_state->complete_count != size) {
opal_progress(); opal_progress();
opal_atomic_mb(); opal_atomic_mb();
} }
OBJ_RELEASE(module->post_group); OBJ_RELEASE(group);
module->post_group = NULL; module->post_group = NULL;
}
OPAL_THREAD_UNLOCK(&module->lock);
/* ensure all memory operations have completed */ /* ensure all memory operations have completed */
opal_atomic_mb(); opal_atomic_mb();
@ -183,7 +302,13 @@ ompi_osc_sm_test(struct ompi_win_t *win,
ompi_osc_sm_module_t *module = ompi_osc_sm_module_t *module =
(ompi_osc_sm_module_t*) win->w_osc_module; (ompi_osc_sm_module_t*) win->w_osc_module;
if (NULL != module->post_group) { OPAL_THREAD_LOCK(&module->lock);
if (NULL == module->post_group) {
OPAL_THREAD_UNLOCK(&module->lock);
return OMPI_ERR_RMA_SYNC;
}
int size = ompi_group_size(module->post_group); int size = ompi_group_size(module->post_group);
if (module->my_node_state->complete_count == size) { if (module->my_node_state->complete_count == size) {
@ -191,10 +316,8 @@ ompi_osc_sm_test(struct ompi_win_t *win,
module->post_group = NULL; module->post_group = NULL;
*flag = 1; *flag = 1;
} }
} else {
opal_atomic_mb(); OPAL_THREAD_UNLOCK(&module->lock);
*flag = 0;
}
/* ensure all memory operations have completed */ /* ensure all memory operations have completed */
opal_atomic_mb(); opal_atomic_mb();

108
ompi/mca/osc/sm/osc_sm_component.c
Просмотреть файл

@ -58,45 +58,45 @@ ompi_osc_sm_component_t mca_osc_sm_component = {
ompi_osc_sm_module_t ompi_osc_sm_module_template = { ompi_osc_sm_module_t ompi_osc_sm_module_template = {
{ {
ompi_osc_sm_shared_query, .osc_win_shared_query = ompi_osc_sm_shared_query,
ompi_osc_sm_attach, .osc_win_attach = ompi_osc_sm_attach,
ompi_osc_sm_detach, .osc_win_detach = ompi_osc_sm_detach,
ompi_osc_sm_free, .osc_free = ompi_osc_sm_free,
ompi_osc_sm_put, .osc_put = ompi_osc_sm_put,
ompi_osc_sm_get, .osc_get = ompi_osc_sm_get,
ompi_osc_sm_accumulate, .osc_accumulate = ompi_osc_sm_accumulate,
ompi_osc_sm_compare_and_swap, .osc_compare_and_swap = ompi_osc_sm_compare_and_swap,
ompi_osc_sm_fetch_and_op, .osc_fetch_and_op = ompi_osc_sm_fetch_and_op,
ompi_osc_sm_get_accumulate, .osc_get_accumulate = ompi_osc_sm_get_accumulate,
ompi_osc_sm_rput, .osc_rput = ompi_osc_sm_rput,
ompi_osc_sm_rget, .osc_rget = ompi_osc_sm_rget,
ompi_osc_sm_raccumulate, .osc_raccumulate = ompi_osc_sm_raccumulate,
ompi_osc_sm_rget_accumulate, .osc_rget_accumulate = ompi_osc_sm_rget_accumulate,
ompi_osc_sm_fence, .osc_fence = ompi_osc_sm_fence,
ompi_osc_sm_start, .osc_start = ompi_osc_sm_start,
ompi_osc_sm_complete, .osc_complete = ompi_osc_sm_complete,
ompi_osc_sm_post, .osc_post = ompi_osc_sm_post,
ompi_osc_sm_wait, .osc_wait = ompi_osc_sm_wait,
ompi_osc_sm_test, .osc_test = ompi_osc_sm_test,
ompi_osc_sm_lock, .osc_lock = ompi_osc_sm_lock,
ompi_osc_sm_unlock, .osc_unlock = ompi_osc_sm_unlock,
ompi_osc_sm_lock_all, .osc_lock_all = ompi_osc_sm_lock_all,
ompi_osc_sm_unlock_all, .osc_unlock_all = ompi_osc_sm_unlock_all,
ompi_osc_sm_sync, .osc_sync = ompi_osc_sm_sync,
ompi_osc_sm_flush, .osc_flush = ompi_osc_sm_flush,
ompi_osc_sm_flush_all, .osc_flush_all = ompi_osc_sm_flush_all,
ompi_osc_sm_flush_local, .osc_flush_local = ompi_osc_sm_flush_local,
ompi_osc_sm_flush_local_all, .osc_flush_local_all = ompi_osc_sm_flush_local_all,
ompi_osc_sm_set_info, .osc_set_info = ompi_osc_sm_set_info,
ompi_osc_sm_get_info .osc_get_info = ompi_osc_sm_get_info
} }
}; };
@ -163,6 +163,7 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
int flavor, int *model) int flavor, int *model)
{ {
ompi_osc_sm_module_t *module = NULL; ompi_osc_sm_module_t *module = NULL;
int comm_size = ompi_comm_size (comm);
int ret = OMPI_ERROR; int ret = OMPI_ERROR;
if (OMPI_SUCCESS != (ret = check_win_ok(comm, flavor))) { if (OMPI_SUCCESS != (ret = check_win_ok(comm, flavor))) {
@ -174,6 +175,8 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
calloc(1, sizeof(ompi_osc_sm_module_t)); calloc(1, sizeof(ompi_osc_sm_module_t));
if (NULL == module) return OMPI_ERR_TEMP_OUT_OF_RESOURCE; if (NULL == module) return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
OBJ_CONSTRUCT(&module->lock, opal_mutex_t);
/* fill in the function pointer part */ /* fill in the function pointer part */
memcpy(module, &ompi_osc_sm_module_template, memcpy(module, &ompi_osc_sm_module_template,
sizeof(ompi_osc_base_module_t)); sizeof(ompi_osc_base_module_t));
@ -185,7 +188,7 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
module->flavor = flavor; module->flavor = flavor;
/* create the segment */ /* create the segment */
if (1 == ompi_comm_size(comm)) { if (1 == comm_size) {
module->segment_base = NULL; module->segment_base = NULL;
module->sizes = malloc(sizeof(size_t)); module->sizes = malloc(sizeof(size_t));
if (NULL == module->sizes) return OMPI_ERR_TEMP_OUT_OF_RESOURCE; if (NULL == module->sizes) return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
@ -200,13 +203,16 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
if (NULL == module->global_state) return OMPI_ERR_TEMP_OUT_OF_RESOURCE; if (NULL == module->global_state) return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
module->node_states = malloc(sizeof(ompi_osc_sm_node_state_t)); module->node_states = malloc(sizeof(ompi_osc_sm_node_state_t));
if (NULL == module->node_states) return OMPI_ERR_TEMP_OUT_OF_RESOURCE; if (NULL == module->node_states) return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
module->posts = calloc (1, sizeof(module->posts[0]) + sizeof (uint64_t));
if (NULL == module->posts) return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
module->posts[0] = (uint64_t *) (module->posts + 1);
} else { } else {
unsigned long total, *rbuf; unsigned long total, *rbuf;
char *data_file; char *data_file;
int i, flag; int i, flag;
size_t pagesize; size_t pagesize;
size_t state_size; size_t state_size;
int posts_size, post_size = (comm_size + 63) / 64;
OPAL_OUTPUT_VERBOSE((1, ompi_osc_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((1, ompi_osc_base_framework.framework_output,
"allocating shared memory region of size %ld\n", (long) size)); "allocating shared memory region of size %ld\n", (long) size));
@ -214,7 +220,7 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
/* get the pagesize */ /* get the pagesize */
pagesize = opal_getpagesize(); pagesize = opal_getpagesize();
rbuf = malloc(sizeof(unsigned long) * ompi_comm_size(module->comm)); rbuf = malloc(sizeof(unsigned long) * comm_size);
if (NULL == rbuf) return OMPI_ERR_TEMP_OUT_OF_RESOURCE; if (NULL == rbuf) return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
module->noncontig = false; module->noncontig = false;
@ -235,7 +241,7 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
if (OMPI_SUCCESS != ret) return ret; if (OMPI_SUCCESS != ret) return ret;
total = 0; total = 0;
for (i = 0 ; i < ompi_comm_size(module->comm) ; ++i) { for (i = 0 ; i < comm_size ; ++i) {
total += rbuf[i]; total += rbuf[i];
} }
@ -247,9 +253,10 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
} }
/* user opal/shmem directly to create a shared memory segment */ /* user opal/shmem directly to create a shared memory segment */
state_size = sizeof(ompi_osc_sm_global_state_t) + sizeof(ompi_osc_sm_node_state_t) * ompi_comm_size(module->comm); state_size = sizeof(ompi_osc_sm_global_state_t) + sizeof(ompi_osc_sm_node_state_t) * comm_size;
posts_size = comm_size * post_size * sizeof (uint64_t);
if (0 == ompi_comm_rank (module->comm)) { if (0 == ompi_comm_rank (module->comm)) {
ret = opal_shmem_segment_create (&module->seg_ds, data_file, total + pagesize + state_size); ret = opal_shmem_segment_create (&module->seg_ds, data_file, total + pagesize + state_size + posts_size);
if (OPAL_SUCCESS != ret) { if (OPAL_SUCCESS != ret) {
goto error; goto error;
} }
@ -266,15 +273,22 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
goto error; goto error;
} }
module->sizes = malloc(sizeof(size_t) * ompi_comm_size(module->comm)); module->sizes = malloc(sizeof(size_t) * comm_size);
if (NULL == module->sizes) return OMPI_ERR_TEMP_OUT_OF_RESOURCE; if (NULL == module->sizes) return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
module->bases = malloc(sizeof(void*) * ompi_comm_size(module->comm)); module->bases = malloc(sizeof(void*) * comm_size);
if (NULL == module->bases) return OMPI_ERR_TEMP_OUT_OF_RESOURCE; if (NULL == module->bases) return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
module->posts = calloc (comm_size, sizeof (module->posts[0]));
if (NULL == module->posts) return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
module->global_state = (ompi_osc_sm_global_state_t *) (module->segment_base); module->global_state = (ompi_osc_sm_global_state_t *) (module->segment_base);
module->node_states = (ompi_osc_sm_node_state_t *) (module->global_state + 1); module->node_states = (ompi_osc_sm_node_state_t *) (module->global_state + 1);
module->posts[0] = (uint64_t *) (module->node_states + comm_size);
for (i = 0, total = state_size + posts_size ; i < comm_size ; ++i) {
if (i > 0) {
module->posts[i] = module->posts[i - 1] + post_size;
}
for (i = 0, total = state_size ; i < ompi_comm_size(module->comm) ; ++i) {
module->sizes[i] = rbuf[i]; module->sizes[i] = rbuf[i];
if (module->sizes[i]) { if (module->sizes[i]) {
module->bases[i] = ((char *) module->segment_base) + total; module->bases[i] = ((char *) module->segment_base) + total;
@ -296,7 +310,7 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
opal_atomic_init(&module->my_node_state->accumulate_lock, OPAL_ATOMIC_UNLOCKED); opal_atomic_init(&module->my_node_state->accumulate_lock, OPAL_ATOMIC_UNLOCKED);
/* share everyone's displacement units. */ /* share everyone's displacement units. */
module->disp_units = malloc(sizeof(int) * ompi_comm_size(module->comm)); module->disp_units = malloc(sizeof(int) * comm_size);
ret = module->comm->c_coll.coll_allgather(&disp_unit, 1, MPI_INT, ret = module->comm->c_coll.coll_allgather(&disp_unit, 1, MPI_INT,
module->disp_units, 1, MPI_INT, module->disp_units, 1, MPI_INT,
module->comm, module->comm,
@ -309,7 +323,7 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
/* initialize synchronization code */ /* initialize synchronization code */
module->my_sense = 1; module->my_sense = 1;
module->outstanding_locks = calloc(ompi_comm_size(module->comm), sizeof(enum ompi_osc_sm_locktype_t)); module->outstanding_locks = calloc(comm_size, sizeof(enum ompi_osc_sm_locktype_t));
if (NULL == module->outstanding_locks) { if (NULL == module->outstanding_locks) {
ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE; ret = OMPI_ERR_TEMP_OUT_OF_RESOURCE;
goto error; goto error;
@ -346,7 +360,7 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
} }
module->global_state->use_barrier_for_fence = 0; module->global_state->use_barrier_for_fence = 0;
module->global_state->sense = module->my_sense; module->global_state->sense = module->my_sense;
module->global_state->count = ompi_comm_size(module->comm); module->global_state->count = comm_size;
pthread_mutexattr_destroy(&mattr); pthread_mutexattr_destroy(&mattr);
} else { } else {
module->global_state->use_barrier_for_fence = 1; module->global_state->use_barrier_for_fence = 1;
@ -367,8 +381,8 @@ component_select(struct ompi_win_t *win, void **base, size_t size, int disp_unit
return OMPI_SUCCESS; return OMPI_SUCCESS;
error: error:
if (NULL != module->comm) ompi_comm_free(&module->comm); win->w_osc_module = &module->super;
if (NULL != module) free(module); ompi_osc_sm_free (win);
return ret; return ret;
} }
@ -459,10 +473,14 @@ ompi_osc_sm_free(struct ompi_win_t *win)
free(module->sizes); free(module->sizes);
} }
free (module->posts);
/* cleanup */ /* cleanup */
ompi_comm_free(&module->comm); ompi_comm_free(&module->comm);
free(module); free(module);
OBJ_DESTRUCT(&module->lock);
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }