1
1

- CM blocking send/recv optimizations

This patch tries to do as little as possible in the PML CM blocking
    send/receive routines.  Basically, avoid creating and filling in an
    entire request object.  An OMPI-level request is still needed, but we
    can create that on the stack instead of going to a free list.

Signed-off-by: Andrew Friedley <andrew.friedley@intel.com>
Signed-off-by: Jithin Jose <jithin.jose@intel.com>
Этот коммит содержится в:
Jithin Jose 2015-04-02 11:37:09 -07:00
родитель 5f19436cd2
Коммит c09582a3ff
2 изменённых файлов: 55 добавлений и 41 удалений

Просмотреть файл

@ -82,6 +82,13 @@ mca_pml_cm_irecv(void *addr,
}
void mca_pml_cm_recv_fast_completion(struct mca_mtl_request_t *mtl_request)
{
// Do nothing!
ompi_request_complete(mtl_request->ompi_req, true);
return;
}
int
mca_pml_cm_recv(void *addr,
size_t count,
@ -92,35 +99,53 @@ mca_pml_cm_recv(void *addr,
ompi_status_public_t * status)
{
int ret;
mca_pml_cm_thin_recv_request_t *recvreq;
ompi_proc_t* ompi_proc;
MCA_PML_CM_THIN_RECV_REQUEST_ALLOC(recvreq);
if( OPAL_UNLIKELY(NULL == recvreq) ) return OMPI_ERR_OUT_OF_RESOURCE;
ompi_proc_t *ompi_proc;
opal_convertor_t convertor;
mca_pml_cm_request_t req;
mca_mtl_request_t *req_mtl =
alloca(sizeof(mca_mtl_request_t) + ompi_mtl->mtl_request_size);
MCA_PML_CM_THIN_RECV_REQUEST_INIT(recvreq,
ompi_proc,
comm,
src,
datatype,
addr,
count);
MCA_PML_CM_THIN_RECV_REQUEST_START(recvreq, comm, tag, src, ret);
req_mtl->ompi_req = &req.req_ompi;
req_mtl->completion_callback = mca_pml_cm_recv_fast_completion;
req.req_pml_type = MCA_PML_CM_REQUEST_RECV_THIN;
req.req_free_called = false;
req.req_ompi.req_complete = false;
req.req_ompi.req_complete_cb = NULL;
req.req_ompi.req_state = OMPI_REQUEST_ACTIVE;
req.req_ompi.req_status.MPI_TAG = OMPI_ANY_TAG;
req.req_ompi.req_status.MPI_ERROR = OMPI_SUCCESS;
req.req_ompi.req_status._cancelled = 0;
if( MPI_ANY_SOURCE == src ) {
ompi_proc = ompi_proc_local_proc;
} else {
ompi_proc = ompi_comm_peer_lookup( comm, src );
}
opal_convertor_copy_and_prepare_for_recv(
ompi_proc->super.proc_convertor,
&(datatype->super),
count,
addr,
0,
&convertor );
ret = OMPI_MTL_CALL(irecv(ompi_mtl,
comm,
src,
tag,
&convertor,
req_mtl));
if( OPAL_UNLIKELY(OMPI_SUCCESS != ret) ) {
/* BWB - XXX - need cleanup of request here */
MCA_PML_CM_THIN_RECV_REQUEST_RETURN(recvreq);
return ret;
}
ompi_request_wait_completion(&recvreq->req_base.req_ompi);
ompi_request_wait_completion(&req.req_ompi);
if (NULL != status) { /* return status */
*status = recvreq->req_base.req_ompi.req_status;
*status = req.req_ompi.req_status;
}
ret = recvreq->req_base.req_ompi.req_status.MPI_ERROR;
ompi_request_free( (ompi_request_t**)&recvreq );
ret = req.req_ompi.req_status.MPI_ERROR;
return ret;
}

Просмотреть файл

@ -150,31 +150,20 @@ mca_pml_cm_send(void *buf,
ompi_request_free( (ompi_request_t**)&sendreq );
} else {
mca_pml_cm_thin_send_request_t *sendreq;
ompi_proc_t * ompi_proc;
MCA_PML_CM_THIN_SEND_REQUEST_ALLOC(sendreq, comm, dst, ompi_proc);
if (OPAL_UNLIKELY(NULL == sendreq)) return OMPI_ERR_OUT_OF_RESOURCE;
MCA_PML_CM_THIN_SEND_REQUEST_INIT(sendreq,
ompi_proc,
comm,
tag,
dst,
datatype,
sendmode,
buf,
count);
MCA_PML_CM_SEND_REQUEST_START_SETUP((&sendreq->req_send));
opal_convertor_t convertor;
ompi_proc_t *ompi_proc = ompi_comm_peer_lookup(comm, dst);
opal_convertor_copy_and_prepare_for_send(
ompi_proc->super.proc_convertor,
&datatype->super, count, buf, 0,
&convertor);
ret = OMPI_MTL_CALL(send(ompi_mtl,
comm,
dst,
tag,
&sendreq->req_send.req_base.req_convertor,
&convertor,
sendmode));
/* Allow a quick path for the request return */
sendreq->req_send.req_base.req_free_called = true;
MCA_PML_CM_THIN_SEND_REQUEST_PML_COMPLETE(sendreq);
}
return ret;