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

Give a boost to MPI_Barrier.

Просмотр исходного кода 
Based on current implementation it is faster to use a blocking
send than the non-blocking version. Switch the exchange function
used in the barrier to use the blocking version combined with
the non-blocking version of the receive.
Этот коммит содержится в:
George Bosilca 2016-06-02 11:45:25 +09:00
родитель 3b68c1f8db
Коммит 223d75595d
1 изменённых файлов: 18 добавлений и 18 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

36
ompi/mca/coll/base/coll_base_barrier.c
Просмотреть файл

@ -41,26 +41,26 @@
* signal a two peer synchronization.
*/
static inline int
ompi_coll_base_sendrecv_zero(int dest, int stag,
ompi_coll_base_sendrecv_zero( int dest, int stag,
int source, int rtag,
MPI_Comm comm)
MPI_Comm comm )
{
int err, line = 0;
ompi_request_t* reqs[2];
ompi_status_public_t statuses[2];
int err, rc, line = 0;
ompi_request_t* reqs[1];
ompi_status_public_t statuses[1];
/* post new irecv */
err = MCA_PML_CALL(irecv( NULL, 0, MPI_BYTE, source, rtag,
comm, &reqs[0]));
comm, &reqs[0] ));
if (err != MPI_SUCCESS) { line = __LINE__; goto error_handler; }
/* send data to children */
err = MCA_PML_CALL(isend( NULL, 0, MPI_BYTE, dest, stag,
MCA_PML_BASE_SEND_STANDARD, comm, &reqs[1]));
if (err != MPI_SUCCESS) { line = __LINE__; goto error_handler; }
rc = MCA_PML_CALL(send( NULL, 0, MPI_BYTE, dest, stag,
MCA_PML_BASE_SEND_STANDARD, comm ));
if (rc != MPI_SUCCESS) { line = __LINE__; err = rc; goto error_handler; }
err = ompi_request_wait_all( 2, reqs, statuses );
err = ompi_request_wait( &reqs[0], &statuses[0] );
if( MPI_ERR_IN_STATUS == err ) { line = __LINE__;
/* As we use wait_all we will get MPI_ERR_IN_STATUS which is not an error
* code that we can propagate up the stack. Instead, look for the real
@ -201,8 +201,8 @@ int ompi_coll_base_barrier_intra_recursivedoubling(struct ompi_communicator_t *c
/* send message to lower ranked node */
remote = rank - adjsize;
err = ompi_coll_base_sendrecv_zero(remote, MCA_COLL_BASE_TAG_BARRIER,
remote, MCA_COLL_BASE_TAG_BARRIER,
comm);
remote, MCA_COLL_BASE_TAG_BARRIER,
comm);
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl;}
} else if (rank < (size - adjsize)) {
@ -226,8 +226,8 @@ int ompi_coll_base_barrier_intra_recursivedoubling(struct ompi_communicator_t *c
/* post receive from the remote node */
err = ompi_coll_base_sendrecv_zero(remote, MCA_COLL_BASE_TAG_BARRIER,
remote, MCA_COLL_BASE_TAG_BARRIER,
comm);
remote, MCA_COLL_BASE_TAG_BARRIER,
comm);
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl;}
}
}
@ -276,8 +276,8 @@ int ompi_coll_base_barrier_intra_bruck(struct ompi_communicator_t *comm,
/* send message to lower ranked node */
err = ompi_coll_base_sendrecv_zero(to, MCA_COLL_BASE_TAG_BARRIER,
from, MCA_COLL_BASE_TAG_BARRIER,
comm);
from, MCA_COLL_BASE_TAG_BARRIER,
comm);
if (err != MPI_SUCCESS) { line = __LINE__; goto err_hndl;}
}
@ -311,8 +311,8 @@ int ompi_coll_base_barrier_intra_two_procs(struct ompi_communicator_t *comm,
remote = (remote + 1) & 0x1;
err = ompi_coll_base_sendrecv_zero(remote, MCA_COLL_BASE_TAG_BARRIER,
remote, MCA_COLL_BASE_TAG_BARRIER,
comm);
remote, MCA_COLL_BASE_TAG_BARRIER,
comm);
return (err);
}