1
1
openmpi/ompi/request/req_wait.c
George Bosilca e8ae5d0047 Performance improvements. Remove all useless functions call (like when we already
know that the request is inactive we don't have to call fini). Remove if's from
the critical path. Change a do in while to make sure we do the minimum in all cases.
All in all it decrease the latency for mvapi by something between 0.15 and 0.20
micro-seconds. But it's just a first step ...

This commit was SVN r8770.
2006-01-20 17:17:13 +00:00

222 строки
6.7 KiB
C

/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2005 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "ompi/include/constants.h"
#include "request/request.h"
int ompi_request_poll_iterations = 20000;
int ompi_request_wait_any(
size_t count,
ompi_request_t ** requests,
int *index,
ompi_status_public_t * status)
{
#if OMPI_ENABLE_PROGRESS_THREADS
int c;
#endif
size_t i=0, num_requests_null_inactive=0;
int rc = OMPI_SUCCESS;
int completed = -1;
ompi_request_t **rptr=NULL;
ompi_request_t *request=NULL;
#if OMPI_ENABLE_PROGRESS_THREADS
/* poll for completion */
opal_atomic_mb();
for (c = 0; completed < 0 && c < ompi_request_poll_iterations; c++) {
rptr = requests;
num_requests_null_inactive = 0;
for (i = 0; i < count; i++, rptr++) {
request = *rptr;
/*
* Check for null or completed persistent request.
* For MPI_REQUEST_NULL, the req_state is always OMPI_REQUEST_INACTIVE
*/
if( request->req_state == OMPI_REQUEST_INACTIVE ) {
num_requests_null_inactive++;
continue;
}
if (true == request->req_complete) {
completed = i;
goto finished;
}
}
if( num_requests_null_inactive == count )
goto finished;
}
#endif
/* give up and sleep until completion */
OPAL_THREAD_LOCK(&ompi_request_lock);
ompi_request_waiting++;
do {
rptr = requests;
num_requests_null_inactive = 0;
for (i = 0; i < count; i++, rptr++) {
request = *rptr;
/*
* Check for null or completed persistent request.
* For MPI_REQUEST_NULL, the req_state is always OMPI_REQUEST_INACTIVE.
*/
if( request->req_state == OMPI_REQUEST_INACTIVE ) {
num_requests_null_inactive++;
continue;
}
if (request->req_complete == true) {
completed = i;
break;
}
}
if(num_requests_null_inactive == count)
break;
if (completed < 0) {
opal_condition_wait(&ompi_request_cond, &ompi_request_lock);
}
} while (completed < 0);
ompi_request_waiting--;
OPAL_THREAD_UNLOCK(&ompi_request_lock);
#if OMPI_ENABLE_PROGRESS_THREADS
finished:
#endif /* OMPI_ENABLE_PROGRESS_THREADS */
if(num_requests_null_inactive == count) {
*index = MPI_UNDEFINED;
if (MPI_STATUS_IGNORE != status) {
*status = ompi_status_empty;
}
} else {
/* return status */
if (MPI_STATUS_IGNORE != status) {
*status = request->req_status;
}
/* return request to pool */
rc = request->req_fini(rptr);
*index = completed;
}
return rc;
}
int ompi_request_wait_all(
size_t count,
ompi_request_t ** requests,
ompi_status_public_t * statuses)
{
size_t completed = 0, i;
ompi_request_t **rptr;
ompi_request_t *request;
rptr = requests;
for (i = 0; i < count; i++) {
request = *rptr++;
if (request->req_complete == true) {
completed++;
}
}
/* if all requests have not completed -- defer acquiring lock
* unless required
*/
if (completed != count) {
/*
* acquire lock and test for completion - if all requests are
* not completed pend on condition variable until a request
* completes
*/
OPAL_THREAD_LOCK(&ompi_request_lock);
ompi_request_waiting++;
#if OMPI_HAVE_THREAD_SUPPORT
/*
* confirm the status of the pending requests. We have to do it before
* taking the condition or otherwise we can miss some requests completion (the
* one that happpens between our initial test and the aquisition of the lock).
*/
rptr = requests;
for( completed = i = 0; i < count; i++ ) {
request = *rptr++;
if (request->req_complete == true) {
completed++;
}
}
#endif /* OMPI_HAVE_THREAD_SUPPORT */
while( completed != count ) {
/* check number of pending requests */
size_t start = ompi_request_completed;
size_t pending = count - completed;
/*
* wait until at least pending requests complete
*/
while (pending > ompi_request_completed - start) {
opal_condition_wait(&ompi_request_cond, &ompi_request_lock);
}
/*
* confirm that all pending operations have completed.
*/
rptr = requests;
for( completed = i = 0; i < count; i++ ) {
request = *rptr++;
if (request->req_complete == true) {
completed++;
}
}
}
ompi_request_waiting--;
OPAL_THREAD_UNLOCK(&ompi_request_lock);
}
if (MPI_STATUSES_IGNORE != statuses) {
/* fill out status and free request if required */
rptr = requests;
for (i = 0; i < count; i++) {
int rc;
request = *rptr;
if(request == MPI_REQUEST_NULL ||
request->req_state == OMPI_REQUEST_INACTIVE) {
statuses[i] = ompi_status_empty;
} else {
statuses[i] = request->req_status;
rc = request->req_fini(rptr);
if (rc != OMPI_SUCCESS)
return rc;
}
rptr++;
}
} else {
/* free request if required */
rptr = requests;
for (i = 0; i < count; i++) {
int rc;
request = *rptr;
if(request == MPI_REQUEST_NULL ||
request->req_state == OMPI_REQUEST_INACTIVE) {
rc = request->req_fini(rptr);
if (rc != OMPI_SUCCESS)
return rc;
}
rptr++;
}
}
return OMPI_SUCCESS;
}