ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
fec973efda
openmpi/ompi/mca/bcol/basesmuma/bcol_basesmuma_lmsg_bcast.h

627 строки
20 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
#ifdef __PORTALS_AVAIL__
#define __PORTALS_ENABLE__
#include <unistd.h>
#include "ompi_config.h"
#include "ompi/constants.h"
#include "ompi/datatype/ompi_datatype.h"
#include "ompi/communicator/communicator.h"
#include "bcol_basesmuma_utils.h"
#include "bcol_basesmuma_portals.h"
#include "bcol_basesmuma.h"
#if 0
struct scatter_allgather_nb_bcast_state_t
{
/* local variables */
uint64_t length;
int my_rank, src, matched;
int *src_list;
int group_size;
int64_t ready_flag;
int pow_2, pow_2_levels;
int src_list_index;
uint64_t fragment_size; /* user buffer size */
/* Input argument variables */
void *my_userbuf;
int64_t sequence_number;
/* Extra source variables */
bool secondary_root;
int partner , extra_partner;
/* Scatter Allgather offsets */
uint64_t local_sg_offset , global_sg_offset , partner_offset ;
/* Portals messaging relevant variables */
ptl_handle_eq_t allgather_eq_h;
ptl_handle_eq_t read_eq;
ptl_event_t allgather_event;
bool msg_posted;
/* OMPI module and component variables */
mca_bcol_basesmuma_component_t *cs;
mca_bcol_basesmuma_module_t *bcol_module;
/* Control structure and payload variables */
volatile mca_bcol_basesmuma_ctl_struct_t **ctl_structs;
volatile mca_bcol_basesmuma_ctl_struct_t *my_ctl_pointer;
volatile mca_bcol_basesmuma_ctl_struct_t *parent_ctl_pointer; /* scatter source */
volatile mca_bcol_basesmuma_ctl_struct_t *extra_partner_ctl_pointer; /* scatter source */
int phase;
};
typedef struct scatter_allgather_nb_bcast_state_t sg_state_t;
#endif
bool blocked_post = false;
#define IS_SG_DATA_READY(peer, my_flag, my_sequence_number) \
(((peer)->sequence_number == (my_sequence_number) && \
(peer)->flags[BCAST_FLAG] >= (my_flag) \
)? true : false )
#define SG_LARGE_MSG_PROBE(src_list, n_src, src_list_index, matched, \
src, data_buffs, data_src_ctl_pointer, \
data_src_lmsg_ctl_pointer, ready_flag, \
sequence_number) \
do { \
int j; \
for( j = 0; j < n_src; j++) { \
if(src_list[j] != -1) { \
data_src_ctl_pointer = data_buffs[src_list[j]].ctl_struct; \
data_src_lmsg_ctl_pointer = (mca_bcol_basesmuma_portal_buf_addr_t*) \
data_buffs[src_list[j]].payload; \
if( IS_SG_DATA_READY(data_src_ctl_pointer,ready_flag,sequence_number)) { \
src = src_list[j]; \
matched = 1; \
src_list_index = j; \
break; \
} \
} \
} \
} while(0)
#define SG_LARGE_MSG_NB_PROBE(src_list, n_src, src_list_index, matched, \
src, ctl_structs, data_src_ctl_pointer, \
ready_flag, sequence_number) \
do { \
int j; \
for( j = 0; j < n_src; j++) { \
if(src_list[j] != -1) { \
data_src_ctl_pointer = ctl_structs[src_list[j]]; \
if( IS_SG_DATA_READY(data_src_ctl_pointer,ready_flag,sequence_number)) { \
src = src_list[j]; \
matched = 1; \
src_list_index = j; \
break; \
} \
} \
} \
} while(0)
static inline __opal_attribute_always_inline__
int wait_for_peers(int my_rank, int npeers, volatile mca_bcol_basesmuma_payload_t *data_buffs,
int flag_value, int sn)
{
int *peers_list = NULL;
int counter = 0, diter = 0;
volatile mca_bcol_basesmuma_header_t *peer_ctl_pointer = NULL;
peers_list = (int *)malloc(sizeof(int) * npeers);
for (diter = 0; diter < npeers; diter++ ){
peers_list[diter] = my_rank ^ (1<<diter);
assert(peers_list[diter] != -1);
}
counter = 0;
while (counter < npeers) {
for (diter = 0; diter < npeers; diter++){
if (-1 != peers_list[diter]) {
peer_ctl_pointer = data_buffs[peers_list[diter]].ctl_struct;
if (IS_SG_DATA_READY(peer_ctl_pointer, flag_value, sn)) {
counter++;
peers_list[diter] = -1;
}
}
}
opal_progress();
}
return 0;
}
static inline __opal_attribute_always_inline__
int wait_for_peers_nb(int my_rank, int npeers,
volatile mca_bcol_basesmuma_ctl_struct_t **ctl_structs,
volatile int flag_value, int sn)
{
int *peers_list = NULL;
int counter = 0, diter = 0;
volatile mca_bcol_basesmuma_ctl_struct_t *peer_ctl_pointer = NULL;
peers_list = (int *)malloc(sizeof(int) * npeers);
for (diter = 0; diter < npeers; diter++ ){
peers_list[diter] = my_rank ^ (1<<diter);
assert(peers_list[diter] != -1);
}
counter = 0;
while (counter < npeers) {
for (diter = 0; diter < npeers; diter++){
if (-1 != peers_list[diter]) {
peer_ctl_pointer = ctl_structs[peers_list[diter]];
if (IS_SG_DATA_READY(peer_ctl_pointer, flag_value, sn)) {
counter++;
peers_list[diter] = -1;
}
}
}
opal_progress();
}
return 0;
}
static inline __opal_attribute_always_inline__
int wait_for_post_complete_nb(int my_rank, int npeers,
volatile mca_bcol_basesmuma_ctl_struct_t **ctl_structs,
int flag_value, int sn)
{
/* int *peers_list = NULL; */
int peers_list[MAX_SM_GROUP_SIZE];
int counter = 0, diter = 0;
volatile mca_bcol_basesmuma_ctl_struct_t *peer_ctl_pointer = NULL;
/* peers_list = (int *)malloc(sizeof(int) * npeers); */
assert(npeers < MAX_SM_GROUP_SIZE);
for (diter = 0; diter < npeers; diter++ ){
peers_list[diter] = my_rank ^ (1<<diter);
assert(peers_list[diter] != -1);
}
counter = 0;
for (diter = 0; diter < npeers; diter++){
peer_ctl_pointer = ctl_structs[peers_list[diter]];
if (IS_SG_DATA_READY(peer_ctl_pointer, flag_value, sn)) {
counter++;
}
}
/* free(peers_list); */
return counter;
}
static inline __opal_attribute_always_inline__
int sg_large_msg_probe(sg_state_t *sg_state)
{
int j,n_src = sg_state->pow_2_levels+1;
for( j = 0; j < n_src; j++) {
if(sg_state->src_list[j] != -1) {
sg_state->parent_ctl_pointer = sg_state->ctl_structs[sg_state->src_list[j]];
BASESMUMA_VERBOSE(5,("Parent %d ctl pointer (parent=%x, my ctl=%x) flag %d",
sg_state->src_list[j],sg_state->parent_ctl_pointer,
sg_state->my_ctl_pointer,
sg_state->parent_ctl_pointer->flag));
if (IS_SG_DATA_READY(sg_state->parent_ctl_pointer,
sg_state->ready_flag, sg_state->sequence_number)) {
sg_state->src = sg_state->src_list[j];
sg_state->matched = 1;
sg_state->src_list_index = j;
break;
}
}
}
return 0;
}
/*
* I will post message for all the my children
*/
static inline __opal_attribute_always_inline__
int sm_portals_root_scatter(sg_state_t *sg_state)
{
int extra_src_posts = -1, scatter_posts = -1, allgather_posts = -1,
total_msg_posts = -1;
BASESMUMA_VERBOSE(10,("I am the root of the data"));
sg_state->my_ctl_pointer->offset = 0;
sg_state->my_ctl_pointer->n_sends = sg_state->pow_2_levels;
sg_state->my_ctl_pointer->length = sg_state->fragment_size;
extra_src_posts = (sg_state->my_rank + sg_state->pow_2 < sg_state->group_size ) ? 1: 0;
scatter_posts = sg_state->my_ctl_pointer->n_sends;
allgather_posts = sg_state->pow_2_levels - 1;
total_msg_posts = scatter_posts + allgather_posts + extra_src_posts ;
if ( total_msg_posts <= 0) {
BASESMUMA_VERBOSE(10,("No need to post the data "));
return OMPI_SUCCESS;
}
mca_bcol_basesmuma_portals_post_msg(sg_state->cs,
&sg_state->my_ctl_pointer->portals_buf_addr,
sg_state->my_userbuf, sg_state->fragment_size,
PTL_EQ_NONE,
total_msg_posts,
blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE |
PTL_MD_OP_GET | PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE);
/*
mca_bcol_basesmuma_portals_post_msg(sg_state->cs,
&sg_state->my_ctl_pointer->portals_buf_addr,
sg_state->my_userbuf, sg_state->fragment_size,
sg_state->allgather_eq_h,
total_msg_posts,
blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE |
PTL_MD_OP_GET | PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE);
*/
sg_state->msg_posted = true ;
/*
basesmuma: missed a couple of MB references cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30538. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 22:19:53 +04:00
opal_atomic_wmb();
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
*/
sg_state->my_ctl_pointer->flag = sg_state->ready_flag;
return OMPI_SUCCESS;
}
/*
* Im root but my rank > pow2_groupsize, so will copy to partner who
* will act as root (secondary)
*/
static inline __opal_attribute_always_inline__
int sm_portals_extra_root_scatter(sg_state_t *sg_state)
{
int scatter_partner = -1;
volatile mca_bcol_basesmuma_ctl_struct_t *scatter_partner_ctl_pointer = NULL;
int total_msg_posts = 1;
if ( total_msg_posts <= 0) {
BASESMUMA_VERBOSE(10,("No need to post the data "));
}
else {
mca_bcol_basesmuma_portals_post_msg(sg_state->cs,
&sg_state->my_ctl_pointer->portals_buf_addr,
sg_state->my_userbuf, sg_state->fragment_size,
PTL_EQ_NONE,
total_msg_posts,
blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE | PTL_MD_OP_GET
| PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE);
sg_state->msg_posted = true ;
}
basesmuma: missed a couple of MB references cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30538. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 22:19:53 +04:00
opal_atomic_wmb();
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
sg_state->my_ctl_pointer->flag = sg_state->ready_flag;
scatter_partner = sg_state->my_rank - sg_state->pow_2;
scatter_partner_ctl_pointer =
sg_state->ctl_structs[scatter_partner];
while(!IS_SG_DATA_READY(scatter_partner_ctl_pointer, sg_state->ready_flag,
sg_state->sequence_number)){
opal_progress();
}
return OMPI_SUCCESS;
}
/*
* Gets msg from the partner (> pow2_groupsize) and posts the
* message acting as root
*/
static inline __opal_attribute_always_inline__
int sm_portals_secondary_root_scatter(sg_state_t *sg_state)
{
volatile mca_bcol_basesmuma_ctl_struct_t *extra_src_ctl_pointer = NULL;
int scatter_posts, allgather_posts, extra_src_posts, total_msg_posts;
sg_state->secondary_root = true;
BASESMUMA_VERBOSE(10,("I am the secondary root for the data"));
sg_state->my_ctl_pointer->offset = 0;
sg_state->my_ctl_pointer->n_sends = sg_state->pow_2_levels;
sg_state->my_ctl_pointer->length = sg_state->fragment_size;
extra_src_ctl_pointer = sg_state->ctl_structs[sg_state->src];
mca_bcol_basesmuma_portals_get_msg_fragment(sg_state->cs,
sg_state->read_eq,
&sg_state->my_ctl_pointer->portals_buf_addr,
&extra_src_ctl_pointer->portals_buf_addr, 0,
0, sg_state->fragment_size);
extra_src_posts = 0;
scatter_posts = sg_state->my_ctl_pointer->n_sends;
allgather_posts = sg_state->pow_2_levels - 1;
total_msg_posts = scatter_posts + allgather_posts + extra_src_posts ;
if (total_msg_posts > 0) {
mca_bcol_basesmuma_portals_post_msg(sg_state->cs,
&sg_state->my_ctl_pointer->portals_buf_addr,
sg_state->my_userbuf, sg_state->fragment_size,
PTL_EQ_NONE,
total_msg_posts,
blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE | PTL_MD_OP_GET
| PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE);
sg_state->msg_posted = true ;
}
basesmuma: missed a couple of MB references cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30538. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 22:19:53 +04:00
opal_atomic_wmb();
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
sg_state->my_ctl_pointer->flag = sg_state->ready_flag;
return OMPI_SUCCESS;
}
/*
* Internode Scatter: Get data from my parent and post for my children
*/
static inline __opal_attribute_always_inline__
int sm_portals_internode_scatter(sg_state_t *sg_state)
{
int scatter_posts, allgather_posts, extra_src_posts,
total_msg_posts;
uint64_t local_offset, remote_offset;
/* compute the size of the chunk to copy */
sg_state->length = (sg_state->parent_ctl_pointer->length)/
(1<<(sg_state->parent_ctl_pointer->n_sends - sg_state->my_ctl_pointer->n_sends));
sg_state->my_ctl_pointer->length = sg_state->length;
sg_state->my_ctl_pointer->offset =
sg_state->parent_ctl_pointer->offset + sg_state->length;
local_offset = sg_state->my_ctl_pointer->offset;
remote_offset = sg_state->parent_ctl_pointer->offset +
sg_state->length;
mca_bcol_basesmuma_portals_get_msg_fragment(sg_state->cs,
sg_state->read_eq,
&sg_state->my_ctl_pointer->portals_buf_addr,
&sg_state->parent_ctl_pointer->portals_buf_addr,local_offset,
remote_offset,sg_state->length);
/* Now post the message for other children to read */
extra_src_posts = (sg_state->my_rank + sg_state->pow_2 <
sg_state->group_size ) ? 1: 0;
scatter_posts = sg_state->my_ctl_pointer->n_sends;
allgather_posts = sg_state->pow_2_levels - 1;
total_msg_posts = scatter_posts + allgather_posts + extra_src_posts ;
if (total_msg_posts > 0) {
mca_bcol_basesmuma_portals_post_msg(sg_state->cs, &sg_state->my_ctl_pointer->portals_buf_addr,
sg_state->my_userbuf, sg_state->my_ctl_pointer->portals_buf_addr.userbuf_length,
PTL_EQ_NONE,
total_msg_posts,
blocked_post,
PTL_MD_EVENT_START_DISABLE| PTL_MD_EVENT_END_DISABLE
| PTL_MD_OP_GET | PTL_MD_MANAGE_REMOTE | PTL_MD_TRUNCATE | PTL_MD_EVENT_AUTO_UNLINK_ENABLE);
sg_state->msg_posted = true;
}
/*
basesmuma: missed a couple of MB references cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30538. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 22:19:53 +04:00
opal_atomic_wmb();
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
*/
sg_state->my_ctl_pointer->flag = sg_state->ready_flag;
return OMPI_SUCCESS;
}
/*
* Bcast's Allgather Phase:
* Combines data from all processes using recursive doubling algorithm
*/
static inline __opal_attribute_always_inline__
int sm_portals_bcasts_allgather_phase(sg_state_t *sg_state)
{
int ag_loop, partner;
volatile mca_bcol_basesmuma_ctl_struct_t *partner_ctl_pointer = NULL; /* recursive double */
for( ag_loop = 1; ag_loop < sg_state->pow_2_levels; ag_loop++) {
/* get my partner for this level */
partner = sg_state->my_rank^(1<<ag_loop);
partner_ctl_pointer = sg_state->ctl_structs[partner];
/* Block until partner is at this level of recursive-doubling stage */
while(!IS_SG_DATA_READY(partner_ctl_pointer, sg_state->ready_flag,
sg_state->sequence_number)) {
opal_progress();
}
assert(partner_ctl_pointer->flag >= sg_state->ready_flag);
if (partner_ctl_pointer->offset < sg_state->my_ctl_pointer->offset) {
sg_state->global_sg_offset -= sg_state->length;
sg_state->local_sg_offset = sg_state->global_sg_offset;
} else {
sg_state->local_sg_offset = sg_state->global_sg_offset + sg_state->length;
}
BASESMUMA_VERBOSE(10,("Allgather Phase: Get message from process %d, length %d",
partner, sg_state->length));
mca_bcol_basesmuma_portals_get_msg_fragment(sg_state->cs,
sg_state->read_eq,
&sg_state->my_ctl_pointer->portals_buf_addr,
&partner_ctl_pointer->portals_buf_addr,sg_state->local_sg_offset,
sg_state->local_sg_offset, sg_state->length);
sg_state->ready_flag++;
basesmuma: missed a couple of MB references cmr=v1.7.5:ticket=trac:4158 This commit was SVN r30538. The following Trac tickets were found above: Ticket 4158 --> https://svn.open-mpi.org/trac/ompi/ticket/4158
2014-02-03 22:19:53 +04:00
opal_atomic_wmb();
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
sg_state->my_ctl_pointer->flag = sg_state->ready_flag;
/* Block until partner is at this level of recursive-doubling stage */
while(!IS_SG_DATA_READY(partner_ctl_pointer, sg_state->ready_flag,
sg_state->sequence_number)) {
opal_progress();
}
/* double the length */
sg_state->length *= 2;
}
return OMPI_SUCCESS;
}
static inline __opal_attribute_always_inline__
int init_sm_group_info(sg_state_t *sg_state, int buff_idx)
{
int idx, leading_dim;
int first_instance=0;
int flag_offset;
/* Get addresing information */
sg_state->group_size = sg_state->bcol_module->colls_no_user_data.size_of_group;
leading_dim = sg_state->bcol_module->colls_no_user_data.size_of_group;
idx=SM_ARRAY_INDEX(leading_dim,buff_idx,0);
BASESMUMA_VERBOSE(1,("My buffer idx %d group size %d, leading dim %d, idx %d",
buff_idx,sg_state->group_size,leading_dim,idx));
/* grab the ctl buffs */
sg_state->ctl_structs = (volatile mca_bcol_basesmuma_ctl_struct_t **)
sg_state->bcol_module->colls_with_user_data.ctl_buffs+idx;
sg_state->my_rank = sg_state->bcol_module->super.sbgp_partner_module->my_index;
sg_state->my_ctl_pointer = sg_state->ctl_structs[sg_state->my_rank];
if (sg_state->my_ctl_pointer->sequence_number < sg_state->sequence_number) {
first_instance = 1;
}
if(first_instance) {
sg_state->my_ctl_pointer->flag = -1;
sg_state->my_ctl_pointer->index = 1;
sg_state->my_ctl_pointer->starting_flag_value = 0;
flag_offset = 0;
} else {
sg_state->my_ctl_pointer->index++;
}
/* For bcast we shud have only entry to this bcol
assert(sg_state->my_ctl_pointer->flag == -1);
*/
/* increment the starting flag by one and return */
flag_offset = sg_state->my_ctl_pointer->starting_flag_value;
sg_state->ready_flag = flag_offset + sg_state->sequence_number + 1;
sg_state->my_ctl_pointer->sequence_number = sg_state->sequence_number;
return OMPI_SUCCESS;
}
static inline __opal_attribute_always_inline__
int init_sm_portals_sg_info(sg_state_t *sg_state)
{
/* Get portals info*/
mca_bcol_basesmuma_portal_proc_info_t *portals_info;
int rc = OMPI_SUCCESS;
int sg_matchbits;
portals_info = (mca_bcol_basesmuma_portal_proc_info_t*)sg_state->cs->portals_info;
sg_matchbits = sg_state->sequence_number ;
/* Construct my portal buffer address and copy to payload buffer */
mca_bcol_basesmuma_construct_portal_address(&sg_state->my_ctl_pointer->portals_buf_addr,
portals_info->portal_id.nid,
portals_info->portal_id.pid,
sg_matchbits,
sg_state->bcol_module->super.sbgp_partner_module->group_comm->c_contextid);
sg_state->my_ctl_pointer->portals_buf_addr.userbuf = sg_state->my_userbuf;
sg_state->my_ctl_pointer->portals_buf_addr.userbuf_length = sg_state->fragment_size;
return OMPI_SUCCESS;
}
static inline __opal_attribute_always_inline__
int compute_src_from_root(int group_root, int my_group_rank, int pow2, int
group_size)
{
int root, relative_rank, src, i;
if (group_root < pow2) {
root = group_root;
} else {
/* the source of the data is extra node,
the real root it represented by some rank from
pow2 group */
root = group_root - pow2;
/* shortcut for the case when my rank is root for the group */
if (my_group_rank == root) {
return group_root;
}
}
relative_rank = (my_group_rank - root) < 0 ? my_group_rank - root + pow2 :
my_group_rank - root;
for (i = 1; i < pow2; i<<=1) {
if (relative_rank & i) {
src = my_group_rank ^ i;
if (src >= pow2)
src -= pow2;
return src;
}
}
return -1;
}
int bcol_basesmuma_lmsg_scatter_allgather_portals_bcast(bcol_function_args_t *input_args,
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_function_t *c_input_args);
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
int bcol_basesmuma_lmsg_scatter_allgather_portals_nb_bcast(bcol_function_args_t *input_args,
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_function_t *c_input_args);
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
int bcol_basesmuma_lmsg_scatter_allgather_portals_nb_knownroot_bcast(bcol_function_args_t *input_args,
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_function_t *c_input_args);
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
#endif
Копировать постоянную ссылку