1
1
openmpi/ompi/mca/pml/ob1/pml_ob1_irecv.c
Nathan Hjelm 2b57f4227e ob1: optimize blocking send and receive paths
Per RFC. There are two optimizations in this commit:

 - Allocate requests for blocking sends and receives on the stack. This
   bypasses the request free list and saves two atomics on the critical path.
   This change improves the small message ping-pong by 50-200ns on both AMD
   and Intel CPUs.

 - For small messages try to use the btl sendi function before intializing a
   send request. If the sendi fails or the btl does not have a sendi function
   silently fallback on the standard send path.

cmr=v1.7.5:reviewer=brbarret

This commit was SVN r30343.
2014-01-21 15:16:21 +00:00

312 строки
11 KiB
C

/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2004-2014 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 (c) 2007-2014 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2010-2012 Oracle and/or its affiliates. All rights reserved.
* Copyright (c) 2011 Sandia National Laboratories. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "ompi/request/request.h"
#include "pml_ob1_recvreq.h"
#include "pml_ob1_recvfrag.h"
#include "ompi/peruse/peruse-internal.h"
#include "ompi/message/message.h"
int mca_pml_ob1_irecv_init(void *addr,
size_t count,
ompi_datatype_t * datatype,
int src,
int tag,
struct ompi_communicator_t *comm,
struct ompi_request_t **request)
{
mca_pml_ob1_recv_request_t *recvreq;
MCA_PML_OB1_RECV_REQUEST_ALLOC(recvreq);
if (NULL == recvreq)
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
MCA_PML_OB1_RECV_REQUEST_INIT(recvreq,
addr,
count, datatype, src, tag, comm, true);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&((recvreq)->req_recv.req_base),
PERUSE_RECV);
*request = (ompi_request_t *) recvreq;
return OMPI_SUCCESS;
}
int mca_pml_ob1_irecv(void *addr,
size_t count,
ompi_datatype_t * datatype,
int src,
int tag,
struct ompi_communicator_t *comm,
struct ompi_request_t **request)
{
mca_pml_ob1_recv_request_t *recvreq;
MCA_PML_OB1_RECV_REQUEST_ALLOC(recvreq);
if (NULL == recvreq)
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
MCA_PML_OB1_RECV_REQUEST_INIT(recvreq,
addr,
count, datatype, src, tag, comm, false);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&((recvreq)->req_recv.req_base),
PERUSE_RECV);
MCA_PML_OB1_RECV_REQUEST_START(recvreq);
*request = (ompi_request_t *) recvreq;
return OMPI_SUCCESS;
}
int mca_pml_ob1_recv(void *addr,
size_t count,
ompi_datatype_t * datatype,
int src,
int tag,
struct ompi_communicator_t *comm,
ompi_status_public_t * status)
{
mca_pml_ob1_recv_request_t recvreq;
int rc;
OBJ_CONSTRUCT(&recvreq, mca_pml_ob1_recv_request_t);
MCA_PML_OB1_RECV_REQUEST_INIT(&recvreq, addr, count, datatype,
src, tag, comm, false);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&(recvreq.req_recv.req_base),
PERUSE_RECV);
MCA_PML_OB1_RECV_REQUEST_START(&recvreq);
ompi_request_wait_completion(&recvreq.req_recv.req_base.req_ompi);
if (NULL != status) { /* return status */
*status = recvreq.req_recv.req_base.req_ompi.req_status;
}
rc = recvreq.req_recv.req_base.req_ompi.req_status.MPI_ERROR;
MCA_PML_BASE_RECV_REQUEST_FINI(&recvreq.req_recv);
OBJ_DESTRUCT(&recvreq);
return rc;
}
int
mca_pml_ob1_imrecv( void *buf,
size_t count,
ompi_datatype_t *datatype,
struct ompi_message_t **message,
struct ompi_request_t **request )
{
mca_pml_ob1_recv_frag_t* frag;
mca_pml_ob1_recv_request_t *recvreq;
mca_pml_ob1_hdr_t *hdr;
int src, tag;
ompi_communicator_t *comm;
mca_pml_ob1_comm_proc_t* proc;
mca_pml_ob1_comm_t* ob1_comm;
uint64_t seq;
/* get the request from the message and the frag from the request
before we overwrite everything */
recvreq = (mca_pml_ob1_recv_request_t*) (*message)->req_ptr;
frag = (mca_pml_ob1_recv_frag_t*) recvreq->req_recv.req_base.req_addr;
src = recvreq->req_recv.req_base.req_ompi.req_status.MPI_SOURCE;
tag = recvreq->req_recv.req_base.req_ompi.req_status.MPI_TAG;
comm = (*message)->comm;
ob1_comm = recvreq->req_recv.req_base.req_comm->c_pml_comm;
seq = recvreq->req_recv.req_base.req_sequence;
/* make the request a recv request again */
/* The old request kept pointers to comm and the char datatype.
We're about to release those, but need to make sure comm
doesn't go out of scope (we don't care about the char datatype
anymore). So retain comm, then release the frag, then reinit
the frag (which will retain comm), then release comm (but the
frag still has it's ref, so it'll stay in scope). Make
sense? */
OBJ_RETAIN(comm);
MCA_PML_BASE_RECV_REQUEST_FINI(&recvreq->req_recv);
recvreq->req_recv.req_base.req_type = MCA_PML_REQUEST_RECV;
MCA_PML_OB1_RECV_REQUEST_INIT(recvreq,
buf,
count, datatype,
src, tag, comm, false);
OBJ_RELEASE(comm);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&((recvreq)->req_recv.req_base),
PERUSE_RECV);
/* init/re-init the request */
recvreq->req_lock = 0;
recvreq->req_pipeline_depth = 0;
recvreq->req_bytes_received = 0;
/* What about req_rdma_cnt ? */
recvreq->req_rdma_idx = 0;
recvreq->req_pending = false;
recvreq->req_ack_sent = false;
MCA_PML_BASE_RECV_START(&recvreq->req_recv.req_base);
/* Note - sequence number already assigned */
recvreq->req_recv.req_base.req_sequence = seq;
proc = &ob1_comm->procs[recvreq->req_recv.req_base.req_peer];
recvreq->req_recv.req_base.req_proc = proc->ompi_proc;
prepare_recv_req_converter(recvreq);
/* we can't go through the match, since we already have the match.
Cheat and do what REQUEST_START does, but without the frag
search */
hdr = (mca_pml_ob1_hdr_t*)frag->segments->seg_addr.pval;
switch(hdr->hdr_common.hdr_type) {
case MCA_PML_OB1_HDR_TYPE_MATCH:
mca_pml_ob1_recv_request_progress_match(recvreq, frag->btl, frag->segments,
frag->num_segments);
break;
case MCA_PML_OB1_HDR_TYPE_RNDV:
mca_pml_ob1_recv_request_progress_rndv(recvreq, frag->btl, frag->segments,
frag->num_segments);
break;
case MCA_PML_OB1_HDR_TYPE_RGET:
mca_pml_ob1_recv_request_progress_rget(recvreq, frag->btl, frag->segments,
frag->num_segments);
break;
default:
assert(0);
}
MCA_PML_OB1_RECV_FRAG_RETURN(frag);
ompi_message_return(*message);
*message = MPI_MESSAGE_NULL;
*request = (ompi_request_t *) recvreq;
return OMPI_SUCCESS;
}
int
mca_pml_ob1_mrecv( void *buf,
size_t count,
ompi_datatype_t *datatype,
struct ompi_message_t **message,
ompi_status_public_t* status )
{
mca_pml_ob1_recv_frag_t* frag;
mca_pml_ob1_recv_request_t *recvreq;
mca_pml_ob1_hdr_t *hdr;
int src, tag, rc;
ompi_communicator_t *comm;
mca_pml_ob1_comm_proc_t* proc;
mca_pml_ob1_comm_t* ob1_comm;
uint64_t seq;
/* get the request from the message and the frag from the request
before we overwrite everything */
comm = (*message)->comm;
recvreq = (mca_pml_ob1_recv_request_t*) (*message)->req_ptr;
frag = (mca_pml_ob1_recv_frag_t*) recvreq->req_recv.req_base.req_addr;
src = recvreq->req_recv.req_base.req_ompi.req_status.MPI_SOURCE;
tag = recvreq->req_recv.req_base.req_ompi.req_status.MPI_TAG;
seq = recvreq->req_recv.req_base.req_sequence;
ob1_comm = recvreq->req_recv.req_base.req_comm->c_pml_comm;
/* make the request a recv request again */
/* The old request kept pointers to comm and the char datatype.
We're about to release those, but need to make sure comm
doesn't go out of scope (we don't care about the char datatype
anymore). So retain comm, then release the frag, then reinit
the frag (which will retain comm), then release comm (but the
frag still has it's ref, so it'll stay in scope). Make
sense? */
OBJ_RETAIN(comm);
MCA_PML_BASE_RECV_REQUEST_FINI(&recvreq->req_recv);
recvreq->req_recv.req_base.req_type = MCA_PML_REQUEST_RECV;
MCA_PML_OB1_RECV_REQUEST_INIT(recvreq,
buf,
count, datatype,
src, tag, comm, false);
OBJ_RELEASE(comm);
PERUSE_TRACE_COMM_EVENT (PERUSE_COMM_REQ_ACTIVATE,
&((recvreq)->req_recv.req_base),
PERUSE_RECV);
/* init/re-init the request */
recvreq->req_lock = 0;
recvreq->req_pipeline_depth = 0;
recvreq->req_bytes_received = 0;
recvreq->req_rdma_cnt = 0;
recvreq->req_rdma_idx = 0;
recvreq->req_pending = false;
MCA_PML_BASE_RECV_START(&recvreq->req_recv.req_base);
/* Note - sequence number already assigned */
recvreq->req_recv.req_base.req_sequence = seq;
proc = &ob1_comm->procs[recvreq->req_recv.req_base.req_peer];
recvreq->req_recv.req_base.req_proc = proc->ompi_proc;
prepare_recv_req_converter(recvreq);
/* we can't go through the match, since we already have the match.
Cheat and do what REQUEST_START does, but without the frag
search */
hdr = (mca_pml_ob1_hdr_t*)frag->segments->seg_addr.pval;
switch(hdr->hdr_common.hdr_type) {
case MCA_PML_OB1_HDR_TYPE_MATCH:
mca_pml_ob1_recv_request_progress_match(recvreq, frag->btl, frag->segments,
frag->num_segments);
break;
case MCA_PML_OB1_HDR_TYPE_RNDV:
mca_pml_ob1_recv_request_progress_rndv(recvreq, frag->btl, frag->segments,
frag->num_segments);
break;
case MCA_PML_OB1_HDR_TYPE_RGET:
mca_pml_ob1_recv_request_progress_rget(recvreq, frag->btl, frag->segments,
frag->num_segments);
break;
default:
assert(0);
}
ompi_message_return(*message);
*message = MPI_MESSAGE_NULL;
ompi_request_wait_completion(&(recvreq->req_recv.req_base.req_ompi));
MCA_PML_OB1_RECV_FRAG_RETURN(frag);
if (NULL != status) { /* return status */
*status = recvreq->req_recv.req_base.req_ompi.req_status;
}
rc = recvreq->req_recv.req_base.req_ompi.req_status.MPI_ERROR;
ompi_request_free( (ompi_request_t**)&recvreq );
return rc;
}