1
1
openmpi/ompi/mca/pml/ob1/pml_ob1_sendreq.c

1276 строки
47 KiB
C
Исходник Обычный вид История

/* -*- 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) 2004-2008 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2004-2008 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) 2008 UT-Battelle, LLC. All rights reserved.
* Copyright (c) 2010 Oracle and/or its affiliates. All rights reserved.
* Copyright (c) 2012 NVIDIA Corporation. All rights reserved.
* Copyright (c) 2012 Los Alamos National Security, LLC. All rights
* reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#include "ompi_config.h"
#include "opal/prefetch.h"
#include "ompi/constants.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/mca/btl/btl.h"
#include "orte/mca/errmgr/errmgr.h"
#include "ompi/mca/mpool/mpool.h"
#include "pml_ob1.h"
#include "pml_ob1_hdr.h"
#include "pml_ob1_sendreq.h"
#include "pml_ob1_rdmafrag.h"
#include "pml_ob1_recvreq.h"
#include "ompi/mca/bml/base/base.h"
#include "ompi/memchecker.h"
#include "ompi/memchecker_rw_check.h"
OBJ_CLASS_INSTANCE(mca_pml_ob1_send_range_t, ompi_free_list_item_t,
NULL, NULL);
void mca_pml_ob1_send_request_process_pending(mca_bml_base_btl_t *bml_btl)
{
int rc, i, s = opal_list_get_size(&mca_pml_ob1.send_pending);
/* advance pending requests */
for(i = 0; i < s; i++) {
mca_pml_ob1_send_pending_t pending_type = MCA_PML_OB1_SEND_PENDING_NONE;
mca_pml_ob1_send_request_t* sendreq;
mca_bml_base_btl_t *send_dst;
sendreq = get_request_from_send_pending(&pending_type);
if(OPAL_UNLIKELY(NULL == sendreq))
break;
switch(pending_type) {
case MCA_PML_OB1_SEND_PENDING_SCHEDULE:
rc = mca_pml_ob1_send_request_schedule_exclusive(sendreq);
if(OMPI_ERR_OUT_OF_RESOURCE == rc) {
return;
}
break;
case MCA_PML_OB1_SEND_PENDING_START:
send_dst = mca_bml_base_btl_array_find(
&sendreq->req_endpoint->btl_eager, bml_btl->btl);
if (NULL == send_dst) {
/* Put request back onto pending list and try next one. */
add_request_to_send_pending(sendreq,
MCA_PML_OB1_SEND_PENDING_START, true);
} else {
MCA_PML_OB1_SEND_REQUEST_RESET(sendreq);
rc = mca_pml_ob1_send_request_start_btl(sendreq, send_dst);
if (OMPI_ERR_OUT_OF_RESOURCE == rc) {
/* No more resources on this btl so prepend to the pending
* list to minimize reordering and give up for now. */
add_request_to_send_pending(sendreq,
MCA_PML_OB1_SEND_PENDING_START, false);
return;
}
}
break;
default:
opal_output(0, "[%s:%d] wrong send request type\n",
__FILE__, __LINE__);
break;
}
}
}
/*
* The free call mark the final stage in a request life-cycle. Starting from this
* point the request is completed at both PML and user level, and can be used
* for others p2p communications. Therefore, in the case of the OB1 PML it should
* be added to the free request list.
*/
static int mca_pml_ob1_send_request_free(struct ompi_request_t** request)
{
mca_pml_ob1_send_request_t* sendreq = *(mca_pml_ob1_send_request_t**)request;
assert( false == sendreq->req_send.req_base.req_free_called );
OPAL_THREAD_LOCK(&ompi_request_lock);
sendreq->req_send.req_base.req_free_called = true;
PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_NOTIFY,
&(sendreq->req_send.req_base), PERUSE_SEND );
if( true == sendreq->req_send.req_base.req_pml_complete ) {
/* make buffer defined when the request is compeleted,
and before releasing the objects. */
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_defined,
sendreq->req_send.req_base.req_addr,
sendreq->req_send.req_base.req_count,
sendreq->req_send.req_base.req_datatype);
memchecker_unreg_mem_rw_check(sendreq->req_send.req_base.req_addr,
sendreq->req_send.req_base.req_count,
sendreq->req_send.req_base.req_datatype);
);
MCA_PML_OB1_SEND_REQUEST_RETURN( sendreq );
}
OPAL_THREAD_UNLOCK(&ompi_request_lock);
*request = MPI_REQUEST_NULL;
return OMPI_SUCCESS;
}
static int mca_pml_ob1_send_request_cancel(struct ompi_request_t* request, int complete)
{
/* we dont cancel send requests by now */
return OMPI_SUCCESS;
}
static void mca_pml_ob1_send_request_construct(mca_pml_ob1_send_request_t* req)
{
req->req_send.req_base.req_type = MCA_PML_REQUEST_SEND;
req->req_send.req_base.req_ompi.req_free = mca_pml_ob1_send_request_free;
req->req_send.req_base.req_ompi.req_cancel = mca_pml_ob1_send_request_cancel;
req->req_rdma_cnt = 0;
req->req_throttle_sends = false;
OBJ_CONSTRUCT(&req->req_send_ranges, opal_list_t);
OBJ_CONSTRUCT(&req->req_send_range_lock, opal_mutex_t);
}
static void mca_pml_ob1_send_request_destruct(mca_pml_ob1_send_request_t* req)
{
OBJ_DESTRUCT(&req->req_send_ranges);
OBJ_DESTRUCT(&req->req_send_range_lock);
}
OBJ_CLASS_INSTANCE( mca_pml_ob1_send_request_t,
mca_pml_base_send_request_t,
mca_pml_ob1_send_request_construct,
mca_pml_ob1_send_request_destruct );
/**
* Completion of a short message - nothing left to schedule.
*/
static inline void
mca_pml_ob1_match_completion_free_request( mca_bml_base_btl_t* bml_btl,
mca_pml_ob1_send_request_t* sendreq )
{
if( sendreq->req_send.req_bytes_packed > 0 ) {
PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_XFER_BEGIN,
&(sendreq->req_send.req_base), PERUSE_SEND );
}
/* signal request completion */
send_request_pml_complete(sendreq);
/* check for pending requests */
MCA_PML_OB1_PROGRESS_PENDING(bml_btl);
}
static void
mca_pml_ob1_match_completion_free( struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* des,
int status )
{
mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)des->des_cbdata;
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) des->des_context;
/* check completion status */
if( OPAL_UNLIKELY(OMPI_SUCCESS != status) ) {
/* TSW - FIX */
opal_output(0, "%s:%d FATAL", __FILE__, __LINE__);
orte_errmgr.abort(-1, NULL);
}
mca_pml_ob1_match_completion_free_request( bml_btl, sendreq );
}
static inline void
mca_pml_ob1_rndv_completion_request( mca_bml_base_btl_t* bml_btl,
mca_pml_ob1_send_request_t* sendreq,
size_t req_bytes_delivered )
{
if( sendreq->req_send.req_bytes_packed > 0 ) {
PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_XFER_BEGIN,
&(sendreq->req_send.req_base), PERUSE_SEND );
}
OPAL_THREAD_ADD_SIZE_T(&sendreq->req_bytes_delivered, req_bytes_delivered);
/* advance the request */
OPAL_THREAD_ADD32(&sendreq->req_state, -1);
send_request_pml_complete_check(sendreq);
/* check for pending requests */
MCA_PML_OB1_PROGRESS_PENDING(bml_btl);
}
/*
* Completion of the first fragment of a long message that
* requires an acknowledgement
*/
static void
mca_pml_ob1_rndv_completion( mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* des,
int status )
{
mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)des->des_cbdata;
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*)des->des_context;
size_t req_bytes_delivered;
/* check completion status */
if( OPAL_UNLIKELY(OMPI_SUCCESS != status) ) {
/* TSW - FIX */
opal_output(0, "%s:%d FATAL", __FILE__, __LINE__);
orte_errmgr.abort(-1, NULL);
}
/* count bytes of user data actually delivered. As the rndv completion only
* happens in one thread, the increase of the req_bytes_delivered does not
* have to be atomic.
*/
req_bytes_delivered = mca_pml_ob1_compute_segment_length (btl->btl_seg_size,
(void *) des->des_src,
des->des_src_cnt,
sizeof(mca_pml_ob1_rendezvous_hdr_t));
mca_pml_ob1_rndv_completion_request( bml_btl, sendreq, req_bytes_delivered );
}
/**
* Completion of a get request.
*/
static void
mca_pml_ob1_rget_completion( mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* des,
int status )
{
mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)des->des_cbdata;
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*)des->des_context;
size_t req_bytes_delivered;
/* count bytes of user data actually delivered and check for request completion */
if (OPAL_LIKELY(OMPI_SUCCESS == status)) {
req_bytes_delivered = mca_pml_ob1_compute_segment_length (btl->btl_seg_size,
(void *) des->des_src,
des->des_src_cnt, 0);
OPAL_THREAD_ADD_SIZE_T(&sendreq->req_bytes_delivered, req_bytes_delivered);
}
sendreq->src_des = NULL;
send_request_pml_complete_check(sendreq);
/* free the descriptor */
mca_bml_base_free(bml_btl, des);
MCA_PML_OB1_PROGRESS_PENDING(bml_btl);
}
/**
* Completion of a control message - return resources.
*/
static void
mca_pml_ob1_send_ctl_completion( mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* des,
int status )
{
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) des->des_context;
/* check for pending requests */
MCA_PML_OB1_PROGRESS_PENDING(bml_btl);
}
/**
* Completion of additional fragments of a large message - may need
* to schedule additional fragments.
*/
static void
mca_pml_ob1_frag_completion( mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* des,
int status )
{
mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)des->des_cbdata;
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) des->des_context;
size_t req_bytes_delivered;
/* check completion status */
if( OPAL_UNLIKELY(OMPI_SUCCESS != status) ) {
/* TSW - FIX */
opal_output(0, "%s:%d FATAL", __FILE__, __LINE__);
orte_errmgr.abort(-1, NULL);
}
/* count bytes of user data actually delivered */
req_bytes_delivered = mca_pml_ob1_compute_segment_length (btl->btl_seg_size,
(void *) des->des_src,
des->des_src_cnt,
sizeof(mca_pml_ob1_frag_hdr_t));
OPAL_THREAD_ADD_SIZE_T(&sendreq->req_pipeline_depth, -1);
OPAL_THREAD_ADD_SIZE_T(&sendreq->req_bytes_delivered, req_bytes_delivered);
if(send_request_pml_complete_check(sendreq) == false) {
mca_pml_ob1_send_request_schedule(sendreq);
}
2005-07-19 01:22:55 +04:00
/* check for pending requests */
MCA_PML_OB1_PROGRESS_PENDING(bml_btl);
}
/**
* Buffer the entire message and mark as complete.
*/
int mca_pml_ob1_send_request_start_buffered(
mca_pml_ob1_send_request_t* sendreq,
mca_bml_base_btl_t* bml_btl,
size_t size)
{
mca_btl_base_descriptor_t* des;
mca_btl_base_segment_t* segment;
mca_pml_ob1_hdr_t* hdr;
struct iovec iov;
unsigned int iov_count;
size_t max_data, req_bytes_delivered;
int rc;
/* allocate descriptor */
mca_bml_base_alloc(bml_btl, &des,
MCA_BTL_NO_ORDER,
sizeof(mca_pml_ob1_rendezvous_hdr_t) + size,
MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP);
if( OPAL_UNLIKELY(NULL == des) ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
segment = des->des_src;
/* pack the data into the BTL supplied buffer */
iov.iov_base = (IOVBASE_TYPE*)((unsigned char*)segment->seg_addr.pval +
sizeof(mca_pml_ob1_rendezvous_hdr_t));
iov.iov_len = size;
iov_count = 1;
max_data = size;
- Split the datatype engine into two parts: an MPI specific part in OMPI and a language agnostic part in OPAL. The convertor is completely moved into OPAL. This offers several benefits as described in RFC http://www.open-mpi.org/community/lists/devel/2009/07/6387.php namely: - Fewer basic types (int* and float* types, boolean and wchar - Fixing naming scheme to ompi-nomenclature. - Usability outside of the ompi-layer. - Due to the fixed nature of simple opal types, their information is completely known at compile time and therefore constified - With fewer datatypes (22), the actual sizes of bit-field types may be reduced from 64 to 32 bits, allowing reorganizing the opal_datatype structure, eliminating holes and keeping data required in convertor (upon send/recv) in one cacheline... This has implications to the convertor-datastructure and other parts of the code. - Several performance tests have been run, the netpipe latency does not change with this patch on Linux/x86-64 on the smoky cluster. - Extensive tests have been done to verify correctness (no new regressions) using: 1. mpi_test_suite on linux/x86-64 using clean ompi-trunk and ompi-ddt: a. running both trunk and ompi-ddt resulted in no differences (except for MPI_SHORT_INT and MPI_TYPE_MIX_LB_UB do now run correctly). b. with --enable-memchecker and running under valgrind (one buglet when run with static found in test-suite, commited) 2. ibm testsuite on linux/x86-64 using clean ompi-trunk and ompi-ddt: all passed (except for the dynamic/ tests failed!! as trunk/MTT) 3. compilation and usage of HDF5 tests on Jaguar using PGI and PathScale compilers. 4. compilation and usage on Scicortex. - Please note, that for the heterogeneous case, (-m32 compiled binaries/ompi), neither ompi-trunk, nor ompi-ddt branch would successfully launch. This commit was SVN r21641.
2009-07-13 08:56:31 +04:00
if((rc = opal_convertor_pack( &sendreq->req_send.req_base.req_convertor,
&iov,
&iov_count,
&max_data)) < 0) {
mca_bml_base_free(bml_btl, des);
return rc;
}
req_bytes_delivered = max_data;
/* build rendezvous header */
hdr = (mca_pml_ob1_hdr_t*)segment->seg_addr.pval;
hdr->hdr_common.hdr_flags = 0;
hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_RNDV;
hdr->hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid;
hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank;
hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag;
hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence;
hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed;
hdr->hdr_rndv.hdr_src_req.pval = sendreq;
ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_RNDV,
sendreq->req_send.req_base.req_proc);
/* update lengths */
segment->seg_len = sizeof(mca_pml_ob1_rendezvous_hdr_t) + max_data;
des->des_cbfunc = mca_pml_ob1_rndv_completion;
des->des_cbdata = sendreq;
/* buffer the remainder of the message */
rc = mca_pml_base_bsend_request_alloc((ompi_request_t*)sendreq);
if( OPAL_UNLIKELY(OMPI_SUCCESS != rc) ) {
mca_bml_base_free(bml_btl, des);
return rc;
}
iov.iov_base = (IOVBASE_TYPE*)(((unsigned char*)sendreq->req_send.req_addr) + max_data);
iov.iov_len = max_data = sendreq->req_send.req_bytes_packed - max_data;
- Split the datatype engine into two parts: an MPI specific part in OMPI and a language agnostic part in OPAL. The convertor is completely moved into OPAL. This offers several benefits as described in RFC http://www.open-mpi.org/community/lists/devel/2009/07/6387.php namely: - Fewer basic types (int* and float* types, boolean and wchar - Fixing naming scheme to ompi-nomenclature. - Usability outside of the ompi-layer. - Due to the fixed nature of simple opal types, their information is completely known at compile time and therefore constified - With fewer datatypes (22), the actual sizes of bit-field types may be reduced from 64 to 32 bits, allowing reorganizing the opal_datatype structure, eliminating holes and keeping data required in convertor (upon send/recv) in one cacheline... This has implications to the convertor-datastructure and other parts of the code. - Several performance tests have been run, the netpipe latency does not change with this patch on Linux/x86-64 on the smoky cluster. - Extensive tests have been done to verify correctness (no new regressions) using: 1. mpi_test_suite on linux/x86-64 using clean ompi-trunk and ompi-ddt: a. running both trunk and ompi-ddt resulted in no differences (except for MPI_SHORT_INT and MPI_TYPE_MIX_LB_UB do now run correctly). b. with --enable-memchecker and running under valgrind (one buglet when run with static found in test-suite, commited) 2. ibm testsuite on linux/x86-64 using clean ompi-trunk and ompi-ddt: all passed (except for the dynamic/ tests failed!! as trunk/MTT) 3. compilation and usage of HDF5 tests on Jaguar using PGI and PathScale compilers. 4. compilation and usage on Scicortex. - Please note, that for the heterogeneous case, (-m32 compiled binaries/ompi), neither ompi-trunk, nor ompi-ddt branch would successfully launch. This commit was SVN r21641.
2009-07-13 08:56:31 +04:00
if((rc = opal_convertor_pack( &sendreq->req_send.req_base.req_convertor,
&iov,
&iov_count,
&max_data)) < 0) {
mca_bml_base_free(bml_btl, des);
return rc;
}
/* re-init convertor for packed data */
- Split the datatype engine into two parts: an MPI specific part in OMPI and a language agnostic part in OPAL. The convertor is completely moved into OPAL. This offers several benefits as described in RFC http://www.open-mpi.org/community/lists/devel/2009/07/6387.php namely: - Fewer basic types (int* and float* types, boolean and wchar - Fixing naming scheme to ompi-nomenclature. - Usability outside of the ompi-layer. - Due to the fixed nature of simple opal types, their information is completely known at compile time and therefore constified - With fewer datatypes (22), the actual sizes of bit-field types may be reduced from 64 to 32 bits, allowing reorganizing the opal_datatype structure, eliminating holes and keeping data required in convertor (upon send/recv) in one cacheline... This has implications to the convertor-datastructure and other parts of the code. - Several performance tests have been run, the netpipe latency does not change with this patch on Linux/x86-64 on the smoky cluster. - Extensive tests have been done to verify correctness (no new regressions) using: 1. mpi_test_suite on linux/x86-64 using clean ompi-trunk and ompi-ddt: a. running both trunk and ompi-ddt resulted in no differences (except for MPI_SHORT_INT and MPI_TYPE_MIX_LB_UB do now run correctly). b. with --enable-memchecker and running under valgrind (one buglet when run with static found in test-suite, commited) 2. ibm testsuite on linux/x86-64 using clean ompi-trunk and ompi-ddt: all passed (except for the dynamic/ tests failed!! as trunk/MTT) 3. compilation and usage of HDF5 tests on Jaguar using PGI and PathScale compilers. 4. compilation and usage on Scicortex. - Please note, that for the heterogeneous case, (-m32 compiled binaries/ompi), neither ompi-trunk, nor ompi-ddt branch would successfully launch. This commit was SVN r21641.
2009-07-13 08:56:31 +04:00
opal_convertor_prepare_for_send( &sendreq->req_send.req_base.req_convertor,
&(ompi_mpi_byte.dt.super),
sendreq->req_send.req_bytes_packed,
sendreq->req_send.req_addr );
/* wait for ack and completion */
sendreq->req_state = 2;
/* request is complete at mpi level */
OPAL_THREAD_LOCK(&ompi_request_lock);
MCA_PML_OB1_SEND_REQUEST_MPI_COMPLETE(sendreq, true);
OPAL_THREAD_UNLOCK(&ompi_request_lock);
/* send */
rc = mca_bml_base_send(bml_btl, des, MCA_PML_OB1_HDR_TYPE_RNDV);
if( OPAL_LIKELY( rc >= 0 ) ) {
if( OPAL_LIKELY( 1 == rc ) ) {
mca_pml_ob1_rndv_completion_request( bml_btl, sendreq, req_bytes_delivered);
}
return OMPI_SUCCESS;
}
mca_bml_base_free(bml_btl, des );
return rc;
}
/**
* We work on a buffered request with a size smaller than the eager size
* or the BTL is not able to send the data IN_PLACE. Request a segment
* that is used for initial hdr and any eager data. This is used only
* from the _START macro.
*/
int mca_pml_ob1_send_request_start_copy( mca_pml_ob1_send_request_t* sendreq,
mca_bml_base_btl_t* bml_btl,
size_t size )
{
mca_btl_base_descriptor_t* des = NULL;
mca_btl_base_segment_t* segment;
mca_pml_ob1_hdr_t* hdr;
struct iovec iov;
unsigned int iov_count;
size_t max_data = size;
int rc;
if(NULL != bml_btl->btl->btl_sendi) {
mca_pml_ob1_match_hdr_t match;
match.hdr_common.hdr_flags = 0;
match.hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_MATCH;
match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid;
match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank;
match.hdr_tag = sendreq->req_send.req_base.req_tag;
match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence;
ob1_hdr_hton(&match, MCA_PML_OB1_HDR_TYPE_MATCH,
sendreq->req_send.req_base.req_proc);
/* try to send immediately */
rc = mca_bml_base_sendi( bml_btl, &sendreq->req_send.req_base.req_convertor,
&match, OMPI_PML_OB1_MATCH_HDR_LEN,
size, MCA_BTL_NO_ORDER,
MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP,
MCA_PML_OB1_HDR_TYPE_MATCH,
&des);
if( OPAL_LIKELY(OMPI_SUCCESS == rc) ) {
/* signal request completion */
send_request_pml_complete(sendreq);
/* check for pending requests */
MCA_PML_OB1_PROGRESS_PENDING(bml_btl);
return OMPI_SUCCESS;
}
/* just in case the btl changed the converter, reset it */
if (size > 0 && NULL != des) {
MCA_PML_OB1_SEND_REQUEST_RESET(sendreq);
}
} else {
/* allocate descriptor */
mca_bml_base_alloc( bml_btl, &des,
MCA_BTL_NO_ORDER,
OMPI_PML_OB1_MATCH_HDR_LEN + size,
MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP);
}
if( OPAL_UNLIKELY(NULL == des) ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
segment = des->des_src;
if(size > 0) {
/* pack the data into the supplied buffer */
iov.iov_base = (IOVBASE_TYPE*)((unsigned char*)segment->seg_addr.pval +
OMPI_PML_OB1_MATCH_HDR_LEN);
iov.iov_len = size;
iov_count = 1;
/*
* Before copy the user buffer, make the target part
* accessible.
*/
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_defined,
sendreq->req_send.req_base.req_addr,
sendreq->req_send.req_base.req_count,
sendreq->req_send.req_base.req_datatype);
memchecker_rw_disable_check();
);
- Split the datatype engine into two parts: an MPI specific part in OMPI and a language agnostic part in OPAL. The convertor is completely moved into OPAL. This offers several benefits as described in RFC http://www.open-mpi.org/community/lists/devel/2009/07/6387.php namely: - Fewer basic types (int* and float* types, boolean and wchar - Fixing naming scheme to ompi-nomenclature. - Usability outside of the ompi-layer. - Due to the fixed nature of simple opal types, their information is completely known at compile time and therefore constified - With fewer datatypes (22), the actual sizes of bit-field types may be reduced from 64 to 32 bits, allowing reorganizing the opal_datatype structure, eliminating holes and keeping data required in convertor (upon send/recv) in one cacheline... This has implications to the convertor-datastructure and other parts of the code. - Several performance tests have been run, the netpipe latency does not change with this patch on Linux/x86-64 on the smoky cluster. - Extensive tests have been done to verify correctness (no new regressions) using: 1. mpi_test_suite on linux/x86-64 using clean ompi-trunk and ompi-ddt: a. running both trunk and ompi-ddt resulted in no differences (except for MPI_SHORT_INT and MPI_TYPE_MIX_LB_UB do now run correctly). b. with --enable-memchecker and running under valgrind (one buglet when run with static found in test-suite, commited) 2. ibm testsuite on linux/x86-64 using clean ompi-trunk and ompi-ddt: all passed (except for the dynamic/ tests failed!! as trunk/MTT) 3. compilation and usage of HDF5 tests on Jaguar using PGI and PathScale compilers. 4. compilation and usage on Scicortex. - Please note, that for the heterogeneous case, (-m32 compiled binaries/ompi), neither ompi-trunk, nor ompi-ddt branch would successfully launch. This commit was SVN r21641.
2009-07-13 08:56:31 +04:00
(void)opal_convertor_pack( &sendreq->req_send.req_base.req_convertor,
&iov, &iov_count, &max_data );
/*
* Packing finished, make the user buffer unaccessable.
*/
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_noaccess,
sendreq->req_send.req_base.req_addr,
sendreq->req_send.req_base.req_count,
sendreq->req_send.req_base.req_datatype);
memchecker_rw_enable_check();
);
}
/* build match header */
hdr = (mca_pml_ob1_hdr_t*)segment->seg_addr.pval;
hdr->hdr_common.hdr_flags = 0;
hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_MATCH;
hdr->hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid;
hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank;
hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag;
hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence;
ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_MATCH,
sendreq->req_send.req_base.req_proc);
/* update lengths */
segment->seg_len = OMPI_PML_OB1_MATCH_HDR_LEN + max_data;
/* short message */
des->des_cbdata = sendreq;
des->des_cbfunc = mca_pml_ob1_match_completion_free;
/* send */
rc = mca_bml_base_send_status(bml_btl, des, MCA_PML_OB1_HDR_TYPE_MATCH);
if( OPAL_LIKELY( rc >= OMPI_SUCCESS ) ) {
if( OPAL_LIKELY( 1 == rc ) ) {
mca_pml_ob1_match_completion_free_request( bml_btl, sendreq );
}
return OMPI_SUCCESS;
}
if (OMPI_ERR_RESOURCE_BUSY == rc) {
/* No more resources. Allow the upper level to queue the send */
rc = OMPI_ERR_OUT_OF_RESOURCE;
}
mca_bml_base_free (bml_btl, des);
return rc;
}
/**
* BTL can send directly from user buffer so allow the BTL
* to prepare the segment list. Start sending a small message.
*/
int mca_pml_ob1_send_request_start_prepare( mca_pml_ob1_send_request_t* sendreq,
mca_bml_base_btl_t* bml_btl,
size_t size )
{
mca_btl_base_descriptor_t* des;
mca_btl_base_segment_t* segment;
mca_pml_ob1_hdr_t* hdr;
int rc;
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_defined,
sendreq->req_send.req_base.req_addr,
sendreq->req_send.req_base.req_count,
sendreq->req_send.req_base.req_datatype);
memchecker_rw_disable_check();
);
/* prepare descriptor */
mca_bml_base_prepare_src( bml_btl,
NULL,
&sendreq->req_send.req_base.req_convertor,
MCA_BTL_NO_ORDER,
OMPI_PML_OB1_MATCH_HDR_LEN,
&size,
MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP,
&des );
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_noaccess,
sendreq->req_send.req_base.req_addr,
sendreq->req_send.req_base.req_count,
sendreq->req_send.req_base.req_datatype);
memchecker_rw_enable_check();
);
if( OPAL_UNLIKELY(NULL == des) ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
segment = des->des_src;
/* build match header */
hdr = (mca_pml_ob1_hdr_t*)segment->seg_addr.pval;
hdr->hdr_common.hdr_flags = 0;
hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_MATCH;
hdr->hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid;
hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank;
hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag;
hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence;
ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_MATCH,
sendreq->req_send.req_base.req_proc);
/* short message */
des->des_cbfunc = mca_pml_ob1_match_completion_free;
des->des_cbdata = sendreq;
/* send */
rc = mca_bml_base_send(bml_btl, des, MCA_PML_OB1_HDR_TYPE_MATCH);
if( OPAL_LIKELY( rc >= 0 ) ) {
if( OPAL_LIKELY( 1 == rc ) ) {
mca_pml_ob1_match_completion_free_request( bml_btl, sendreq );
}
return OMPI_SUCCESS;
}
mca_bml_base_free(bml_btl, des );
return rc;
}
/**
* We have contigous data that is registered - schedule across
* available nics.
*/
int mca_pml_ob1_send_request_start_rdma( mca_pml_ob1_send_request_t* sendreq,
mca_bml_base_btl_t* bml_btl,
size_t size )
{
/*
* When req_rdma array is constructed the first element of the array always
* assigned different btl in round robin fashion (if there are more than
* one RDMA capable BTLs). This way round robin distribution of RDMA
* operation is achieved.
*/
mca_btl_base_descriptor_t *des, *src = NULL;
mca_pml_ob1_rget_hdr_t *hdr;
size_t seg_size;
int rc;
sendreq->src_des = NULL;
bml_btl = sendreq->req_rdma[0].bml_btl;
if (!(bml_btl->btl_flags & (MCA_BTL_FLAGS_GET | MCA_BTL_FLAGS_CUDA_GET))) {
/* This BTL does not support get. Use rendezvous to start the RDMA operation using put instead. */
return mca_pml_ob1_send_request_start_rndv (sendreq, bml_btl, 0, MCA_PML_OB1_HDR_FLAGS_CONTIG |
MCA_PML_OB1_HDR_FLAGS_PIN);
}
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_defined,
sendreq->req_send.req_base.req_addr,
sendreq->req_send.req_base.req_count,
sendreq->req_send.req_base.req_datatype);
memchecker_rw_disable_check();
);
/* prepare source descriptor/segment(s) */
/* PML owns this descriptor and will free it in */
/* mca_pml_ob1_rget_completion */
mca_bml_base_prepare_src( bml_btl, sendreq->req_rdma[0].btl_reg,
&sendreq->req_send.req_base.req_convertor,
MCA_BTL_NO_ORDER, 0, &size, MCA_BTL_DES_FLAGS_GET |
MCA_BTL_DES_FLAGS_BTL_OWNERSHIP, &src );
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_noaccess,
sendreq->req_send.req_base.req_addr,
sendreq->req_send.req_base.req_count,
sendreq->req_send.req_base.req_datatype);
memchecker_rw_enable_check();
);
if( OPAL_UNLIKELY(NULL == src) ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
src->des_cbfunc = mca_pml_ob1_rget_completion;
src->des_cbdata = sendreq;
sendreq->src_des = src;
seg_size = bml_btl->btl->btl_seg_size * src->des_src_cnt;
/* allocate space for get hdr + segment list */
mca_bml_base_alloc(bml_btl, &des, MCA_BTL_NO_ORDER, sizeof (*hdr) + seg_size,
MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP);
if( OPAL_UNLIKELY(NULL == des) ) {
/* NTH: no need to reset the converter here. it will be reset before it is retried */
mca_bml_base_free(bml_btl, src);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* build match header */
hdr = (mca_pml_ob1_rget_hdr_t *) des->des_src->seg_addr.pval;
hdr->hdr_rndv.hdr_match.hdr_common.hdr_flags = MCA_PML_OB1_HDR_FLAGS_CONTIG|MCA_PML_OB1_HDR_FLAGS_PIN;
hdr->hdr_rndv.hdr_match.hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_RGET;
hdr->hdr_rndv.hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid;
hdr->hdr_rndv.hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank;
hdr->hdr_rndv.hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag;
hdr->hdr_rndv.hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence;
hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed;
hdr->hdr_rndv.hdr_src_req.pval = sendreq;
hdr->hdr_des.pval = src;
hdr->hdr_seg_cnt = src->des_src_cnt;
ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_RGET, sendreq->req_send.req_base.req_proc);
/* copy segment data */
memcpy (hdr + 1, src->des_src, seg_size);
des->des_cbfunc = mca_pml_ob1_send_ctl_completion;
des->des_cbdata = sendreq;
/**
* Well, it's a get so we will not know when the peer will get the data anyway.
* If we generate the PERUSE event here, at least we will know when we
* sent the GET message ...
*/
if( sendreq->req_send.req_bytes_packed > 0 ) {
PERUSE_TRACE_COMM_EVENT( PERUSE_COMM_REQ_XFER_BEGIN,
&(sendreq->req_send.req_base), PERUSE_SEND );
}
/* send */
rc = mca_bml_base_send(bml_btl, des, MCA_PML_OB1_HDR_TYPE_RGET);
if (OPAL_UNLIKELY(rc < 0)) {
mca_bml_base_free(bml_btl, des);
if (sendreq->src_des) {
mca_bml_base_free (bml_btl, sendreq->src_des);
sendreq->src_des = NULL;
}
return rc;
}
return OMPI_SUCCESS;
}
/**
* Rendezvous is required. Not doing rdma so eager send up to
* the btls eager limit.
*/
int mca_pml_ob1_send_request_start_rndv( mca_pml_ob1_send_request_t* sendreq,
mca_bml_base_btl_t* bml_btl,
size_t size,
int flags )
{
mca_btl_base_descriptor_t* des;
mca_btl_base_segment_t* segment;
mca_pml_ob1_hdr_t* hdr;
int rc;
/* prepare descriptor */
if(size == 0) {
mca_bml_base_alloc( bml_btl,
&des,
MCA_BTL_NO_ORDER,
sizeof(mca_pml_ob1_rendezvous_hdr_t),
MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP );
} else {
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_defined,
sendreq->req_send.req_base.req_addr,
sendreq->req_send.req_base.req_count,
sendreq->req_send.req_base.req_datatype);
memchecker_rw_disable_check();
);
mca_bml_base_prepare_src( bml_btl,
NULL,
&sendreq->req_send.req_base.req_convertor,
MCA_BTL_NO_ORDER,
sizeof(mca_pml_ob1_rendezvous_hdr_t),
&size,
MCA_BTL_DES_FLAGS_PRIORITY | MCA_BTL_DES_FLAGS_BTL_OWNERSHIP,
&des );
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_noaccess,
sendreq->req_send.req_base.req_addr,
sendreq->req_send.req_base.req_count,
sendreq->req_send.req_base.req_datatype);
memchecker_rw_enable_check();
);
}
if( OPAL_UNLIKELY(NULL == des) ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
segment = des->des_src;
/* build hdr */
hdr = (mca_pml_ob1_hdr_t*)segment->seg_addr.pval;
hdr->hdr_common.hdr_flags = flags;
hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_RNDV;
hdr->hdr_match.hdr_ctx = sendreq->req_send.req_base.req_comm->c_contextid;
hdr->hdr_match.hdr_src = sendreq->req_send.req_base.req_comm->c_my_rank;
hdr->hdr_match.hdr_tag = sendreq->req_send.req_base.req_tag;
hdr->hdr_match.hdr_seq = (uint16_t)sendreq->req_send.req_base.req_sequence;
hdr->hdr_rndv.hdr_msg_length = sendreq->req_send.req_bytes_packed;
hdr->hdr_rndv.hdr_src_req.pval = sendreq;
ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_RNDV,
sendreq->req_send.req_base.req_proc);
/* first fragment of a long message */
des->des_cbdata = sendreq;
des->des_cbfunc = mca_pml_ob1_rndv_completion;
/* wait for ack and completion */
sendreq->req_state = 2;
/* send */
rc = mca_bml_base_send(bml_btl, des, MCA_PML_OB1_HDR_TYPE_RNDV);
if( OPAL_LIKELY( rc >= 0 ) ) {
if( OPAL_LIKELY( 1 == rc ) ) {
mca_pml_ob1_rndv_completion_request( bml_btl, sendreq, size );
}
return OMPI_SUCCESS;
}
mca_bml_base_free(bml_btl, des );
return rc;
}
void mca_pml_ob1_send_request_copy_in_out( mca_pml_ob1_send_request_t *sendreq,
uint64_t send_offset,
uint64_t send_length )
{
mca_pml_ob1_send_range_t *sr;
ompi_free_list_item_t *i;
mca_bml_base_endpoint_t* bml_endpoint = sendreq->req_endpoint;
int num_btls = mca_bml_base_btl_array_get_size(&bml_endpoint->btl_send);
int rc, n;
double weight_total = 0;
if( OPAL_UNLIKELY(0 == send_length) )
return;
OMPI_FREE_LIST_WAIT(&mca_pml_ob1.send_ranges, i, rc);
sr = (mca_pml_ob1_send_range_t*)i;
sr->range_send_offset = send_offset;
sr->range_send_length = send_length;
sr->range_btl_idx = 0;
for(n = 0; n < num_btls && n < mca_pml_ob1.max_send_per_range; n++) {
sr->range_btls[n].bml_btl =
mca_bml_base_btl_array_get_next(&bml_endpoint->btl_send);
weight_total += sr->range_btls[n].bml_btl->btl_weight;
}
sr->range_btl_cnt = n;
mca_pml_ob1_calc_weighted_length(sr->range_btls, n, send_length,
weight_total);
OPAL_THREAD_LOCK(&sendreq->req_send_range_lock);
opal_list_append(&sendreq->req_send_ranges, (opal_list_item_t*)sr);
OPAL_THREAD_UNLOCK(&sendreq->req_send_range_lock);
}
static inline mca_pml_ob1_send_range_t *
get_send_range_nolock(mca_pml_ob1_send_request_t* sendreq)
{
opal_list_item_t *item;
item = opal_list_get_first(&sendreq->req_send_ranges);
if(opal_list_get_end(&sendreq->req_send_ranges) == item)
return NULL;
return (mca_pml_ob1_send_range_t*)item;
}
static inline mca_pml_ob1_send_range_t *
get_send_range(mca_pml_ob1_send_request_t* sendreq)
{
mca_pml_ob1_send_range_t *range;
OPAL_THREAD_LOCK(&sendreq->req_send_range_lock);
range = get_send_range_nolock(sendreq);
OPAL_THREAD_UNLOCK(&sendreq->req_send_range_lock);
return range;
}
static inline mca_pml_ob1_send_range_t *
get_next_send_range(mca_pml_ob1_send_request_t* sendreq,
mca_pml_ob1_send_range_t *range)
{
OPAL_THREAD_LOCK(&sendreq->req_send_range_lock);
opal_list_remove_item(&sendreq->req_send_ranges, (opal_list_item_t *)range);
OMPI_FREE_LIST_RETURN(&mca_pml_ob1.send_ranges, &range->base);
range = get_send_range_nolock(sendreq);
OPAL_THREAD_UNLOCK(&sendreq->req_send_range_lock);
return range;
}
/**
* Schedule pipeline of send descriptors for the given request.
* Up to the rdma threshold. If this is a send based protocol,
* the rdma threshold is the end of the message. Otherwise, schedule
* fragments up to the threshold to overlap initial registration/setup
* costs of the rdma. Only one thread can be inside this function.
*/
int
mca_pml_ob1_send_request_schedule_once(mca_pml_ob1_send_request_t* sendreq)
{
size_t prev_bytes_remaining = 0;
mca_pml_ob1_send_range_t *range;
int num_fail = 0;
/* check pipeline_depth here before attempting to get any locks */
if(true == sendreq->req_throttle_sends &&
sendreq->req_pipeline_depth >= mca_pml_ob1.send_pipeline_depth)
return OMPI_SUCCESS;
range = get_send_range(sendreq);
while(range && (false == sendreq->req_throttle_sends ||
sendreq->req_pipeline_depth < mca_pml_ob1.send_pipeline_depth)) {
mca_pml_ob1_frag_hdr_t* hdr;
mca_btl_base_descriptor_t* des;
int rc, btl_idx;
size_t size, offset, data_remaining = 0;
mca_bml_base_btl_t* bml_btl;
assert(range->range_send_length != 0);
if(prev_bytes_remaining == range->range_send_length)
num_fail++;
else
num_fail = 0;
prev_bytes_remaining = range->range_send_length;
if( OPAL_UNLIKELY(num_fail == range->range_btl_cnt) ) {
assert(sendreq->req_pending == MCA_PML_OB1_SEND_PENDING_NONE);
add_request_to_send_pending(sendreq,
MCA_PML_OB1_SEND_PENDING_SCHEDULE, true);
/* Note that request remains locked. send_request_process_pending()
* function will call shedule_exclusive() directly without taking
* the lock */
return OMPI_ERR_OUT_OF_RESOURCE;
}
cannot_pack:
do {
btl_idx = range->range_btl_idx;
if(++range->range_btl_idx == range->range_btl_cnt)
range->range_btl_idx = 0;
} while(!range->range_btls[btl_idx].length);
bml_btl = range->range_btls[btl_idx].bml_btl;
/* If there is a remaining data from another BTL that was too small
* for converter to pack then send it through another BTL */
range->range_btls[btl_idx].length += data_remaining;
size = range->range_btls[btl_idx].length;
/* makes sure that we don't exceed BTL max send size */
if(bml_btl->btl->btl_max_send_size != 0) {
size_t max_send_size = bml_btl->btl->btl_max_send_size -
sizeof(mca_pml_ob1_frag_hdr_t);
if (size > max_send_size) {
size = max_send_size;
}
}
/* pack into a descriptor */
offset = (size_t)range->range_send_offset;
- Split the datatype engine into two parts: an MPI specific part in OMPI and a language agnostic part in OPAL. The convertor is completely moved into OPAL. This offers several benefits as described in RFC http://www.open-mpi.org/community/lists/devel/2009/07/6387.php namely: - Fewer basic types (int* and float* types, boolean and wchar - Fixing naming scheme to ompi-nomenclature. - Usability outside of the ompi-layer. - Due to the fixed nature of simple opal types, their information is completely known at compile time and therefore constified - With fewer datatypes (22), the actual sizes of bit-field types may be reduced from 64 to 32 bits, allowing reorganizing the opal_datatype structure, eliminating holes and keeping data required in convertor (upon send/recv) in one cacheline... This has implications to the convertor-datastructure and other parts of the code. - Several performance tests have been run, the netpipe latency does not change with this patch on Linux/x86-64 on the smoky cluster. - Extensive tests have been done to verify correctness (no new regressions) using: 1. mpi_test_suite on linux/x86-64 using clean ompi-trunk and ompi-ddt: a. running both trunk and ompi-ddt resulted in no differences (except for MPI_SHORT_INT and MPI_TYPE_MIX_LB_UB do now run correctly). b. with --enable-memchecker and running under valgrind (one buglet when run with static found in test-suite, commited) 2. ibm testsuite on linux/x86-64 using clean ompi-trunk and ompi-ddt: all passed (except for the dynamic/ tests failed!! as trunk/MTT) 3. compilation and usage of HDF5 tests on Jaguar using PGI and PathScale compilers. 4. compilation and usage on Scicortex. - Please note, that for the heterogeneous case, (-m32 compiled binaries/ompi), neither ompi-trunk, nor ompi-ddt branch would successfully launch. This commit was SVN r21641.
2009-07-13 08:56:31 +04:00
opal_convertor_set_position(&sendreq->req_send.req_base.req_convertor,
&offset);
range->range_send_offset = (uint64_t)offset;
data_remaining = size;
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_defined,
sendreq->req_send.req_base.req_addr,
sendreq->req_send.req_base.req_count,
sendreq->req_send.req_base.req_datatype);
);
mca_bml_base_prepare_src(bml_btl, NULL,
&sendreq->req_send.req_base.req_convertor,
MCA_BTL_NO_ORDER,
sizeof(mca_pml_ob1_frag_hdr_t),
&size, MCA_BTL_DES_FLAGS_BTL_OWNERSHIP | MCA_BTL_DES_SEND_ALWAYS_CALLBACK, &des);
MEMCHECKER(
memchecker_call(&opal_memchecker_base_mem_noaccess,
sendreq->req_send.req_base.req_addr,
sendreq->req_send.req_base.req_count,
sendreq->req_send.req_base.req_datatype);
);
if( OPAL_UNLIKELY(des == NULL || size == 0) ) {
if(des) {
/* Converter can't pack this chunk. Append to another chunk
* from other BTL */
mca_bml_base_free(bml_btl, des);
range->range_btls[btl_idx].length -= data_remaining;
goto cannot_pack;
}
continue;
}
des->des_cbfunc = mca_pml_ob1_frag_completion;
des->des_cbdata = sendreq;
/* setup header */
hdr = (mca_pml_ob1_frag_hdr_t*)des->des_src->seg_addr.pval;
hdr->hdr_common.hdr_flags = 0;
hdr->hdr_common.hdr_type = MCA_PML_OB1_HDR_TYPE_FRAG;
hdr->hdr_frag_offset = range->range_send_offset;
hdr->hdr_src_req.pval = sendreq;
hdr->hdr_dst_req = sendreq->req_recv;
ob1_hdr_hton(hdr, MCA_PML_OB1_HDR_TYPE_FRAG,
sendreq->req_send.req_base.req_proc);
#if OMPI_WANT_PERUSE
PERUSE_TRACE_COMM_OMPI_EVENT(PERUSE_COMM_REQ_XFER_CONTINUE,
&(sendreq->req_send.req_base), size, PERUSE_SEND);
#endif /* OMPI_WANT_PERUSE */
/* initiate send - note that this may complete before the call returns */
rc = mca_bml_base_send(bml_btl, des, MCA_PML_OB1_HDR_TYPE_FRAG);
if( OPAL_LIKELY(rc >= 0) ) {
/* update state */
range->range_btls[btl_idx].length -= size;
range->range_send_length -= size;
range->range_send_offset += size;
OPAL_THREAD_ADD_SIZE_T(&sendreq->req_pipeline_depth, 1);
if(range->range_send_length == 0) {
range = get_next_send_range(sendreq, range);
prev_bytes_remaining = 0;
}
} else {
mca_bml_base_free(bml_btl,des);
}
}
return OMPI_SUCCESS;
}
/**
* An RDMA put operation has completed:
* (1) Update request status and if required set completed
* (2) Send FIN control message to the destination
*/
static void mca_pml_ob1_put_completion( mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* ep,
struct mca_btl_base_descriptor_t* des,
int status )
{
mca_pml_ob1_rdma_frag_t* frag = (mca_pml_ob1_rdma_frag_t*)des->des_cbdata;
mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)frag->rdma_req;
mca_bml_base_btl_t* bml_btl = (mca_bml_base_btl_t*) des->des_context;
/* check completion status */
if( OPAL_UNLIKELY(OMPI_SUCCESS != status) ) {
/* TSW - FIX */
ORTE_ERROR_LOG(status);
orte_errmgr.abort(-1, NULL);
}
mca_pml_ob1_send_fin(sendreq->req_send.req_base.req_proc,
bml_btl,
frag->rdma_hdr.hdr_rdma.hdr_des,
des->order, 0);
/* check for request completion */
OPAL_THREAD_ADD_SIZE_T(&sendreq->req_bytes_delivered, frag->rdma_length);
send_request_pml_complete_check(sendreq);
MCA_PML_OB1_RDMA_FRAG_RETURN(frag);
MCA_PML_OB1_PROGRESS_PENDING(bml_btl);
}
int mca_pml_ob1_send_request_put_frag( mca_pml_ob1_rdma_frag_t *frag )
{
mca_pml_ob1_send_request_t* sendreq = (mca_pml_ob1_send_request_t*)frag->rdma_req;
mca_mpool_base_registration_t *reg = NULL;
mca_bml_base_btl_t *bml_btl = frag->rdma_bml;
mca_btl_base_descriptor_t *des;
size_t save_size = frag->rdma_length;
int rc;
if (OPAL_LIKELY(NULL == sendreq->src_des)) {
/* setup descriptor */
mca_bml_base_prepare_src( bml_btl,
reg,
&frag->convertor,
MCA_BTL_NO_ORDER,
0,
&frag->rdma_length,
MCA_BTL_DES_FLAGS_BTL_OWNERSHIP |
MCA_BTL_DES_FLAGS_PUT,
&des );
if( OPAL_UNLIKELY(NULL == des) ) {
if(frag->retries < mca_pml_ob1.rdma_retries_limit) {
size_t offset = (size_t)frag->rdma_hdr.hdr_rdma.hdr_rdma_offset;
frag->rdma_length = save_size;
opal_convertor_set_position(&frag->convertor, &offset);
OPAL_THREAD_LOCK(&mca_pml_ob1.lock);
opal_list_append(&mca_pml_ob1.rdma_pending, (opal_list_item_t*)frag);
OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock);
} else {
mca_pml_ob1_send_request_t *sendreq =
(mca_pml_ob1_send_request_t*)frag->rdma_req;
/* tell receiver to unregister memory */
mca_pml_ob1_send_fin(sendreq->req_send.req_base.req_proc,
bml_btl, frag->rdma_hdr.hdr_rdma.hdr_des,
MCA_BTL_NO_ORDER, 1);
/* send fragment by copy in/out */
mca_pml_ob1_send_request_copy_in_out(sendreq,
frag->rdma_hdr.hdr_rdma.hdr_rdma_offset, frag->rdma_length);
/* if a pointer to a receive request is not set it means that
* ACK was not yet received. Don't schedule sends before ACK */
if(NULL != sendreq->req_recv.pval)
mca_pml_ob1_send_request_schedule(sendreq);
}
return OMPI_ERR_OUT_OF_RESOURCE;
}
} else {
/* already have a source descriptor */
des = sendreq->src_des;
sendreq->src_des = NULL;
}
des->des_dst = (mca_btl_base_segment_t *) frag->rdma_segs;
des->des_dst_cnt = frag->rdma_hdr.hdr_rdma.hdr_seg_cnt;
des->des_cbfunc = mca_pml_ob1_put_completion;
des->des_cbdata = frag;
PERUSE_TRACE_COMM_OMPI_EVENT( PERUSE_COMM_REQ_XFER_CONTINUE,
&(((mca_pml_ob1_send_request_t*)frag->rdma_req)->req_send.req_base), save_size, PERUSE_SEND );
rc = mca_bml_base_put(bml_btl, des);
if (OPAL_UNLIKELY(OMPI_SUCCESS != rc)) {
mca_bml_base_free(bml_btl, des);
frag->rdma_length = save_size;
if(OMPI_ERR_OUT_OF_RESOURCE == rc) {
OPAL_THREAD_LOCK(&mca_pml_ob1.lock);
opal_list_append(&mca_pml_ob1.rdma_pending, (opal_list_item_t*)frag);
OPAL_THREAD_UNLOCK(&mca_pml_ob1.lock);
return OMPI_ERR_OUT_OF_RESOURCE;
} else {
/* TSW - FIX */
ORTE_ERROR_LOG(rc);
orte_errmgr.abort(-1, NULL);
}
}
return OMPI_SUCCESS;
}
/**
* Receiver has scheduled an RDMA operation:
* (1) Allocate an RDMA fragment to maintain the state of the operation
* (2) Call BTL prepare_src to pin/prepare source buffers
* (3) Queue the RDMA put
*/
void mca_pml_ob1_send_request_put( mca_pml_ob1_send_request_t* sendreq,
mca_btl_base_module_t* btl,
mca_pml_ob1_rdma_hdr_t* hdr )
{
mca_bml_base_endpoint_t *bml_endpoint = sendreq->req_endpoint;
mca_pml_ob1_rdma_frag_t* frag;
int rc;
size_t i, size = 0;
if(hdr->hdr_common.hdr_flags & MCA_PML_OB1_HDR_TYPE_ACK) {
OPAL_THREAD_ADD32(&sendreq->req_state, -1);
}
sendreq->req_recv.pval = hdr->hdr_recv_req.pval;
MCA_PML_OB1_RDMA_FRAG_ALLOC(frag, rc);
if( OPAL_UNLIKELY(NULL == frag) ) {
/* TSW - FIX */
ORTE_ERROR_LOG(rc);
orte_errmgr.abort(-1, NULL);
}
assert (btl->btl_seg_size * hdr->hdr_seg_cnt <= sizeof (frag->rdma_segs));
/* setup fragment */
memcpy (frag->rdma_segs, hdr + 1, btl->btl_seg_size * hdr->hdr_seg_cnt);
for( i = 0; i < hdr->hdr_seg_cnt; i++ ) {
mca_btl_base_segment_t *seg = (mca_btl_base_segment_t *) ((uintptr_t)(frag->rdma_segs) + i * btl->btl_seg_size);
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
if ((sendreq->req_send.req_base.req_proc->proc_arch & OPAL_ARCH_ISBIGENDIAN) !=
(ompi_proc_local()->proc_arch & OPAL_ARCH_ISBIGENDIAN)) {
size += opal_swap_bytes4(seg->seg_len);
} else
#endif
{
size += seg->seg_len;
}
}
frag->rdma_bml = mca_bml_base_btl_array_find(&bml_endpoint->btl_rdma, btl);
frag->rdma_hdr.hdr_rdma = *hdr;
frag->rdma_req = sendreq;
frag->rdma_ep = bml_endpoint;
frag->rdma_length = size;
frag->rdma_state = MCA_PML_OB1_RDMA_PUT;
frag->reg = NULL;
frag->retries = 0;
if (OPAL_UNLIKELY(NULL != sendreq->src_des)) {
/* get fallback path */
sendreq->req_state = 0;
}
/* lookup the corresponding registration */
for(i=0; i<sendreq->req_rdma_cnt; i++) {
if(sendreq->req_rdma[i].bml_btl == frag->rdma_bml) {
frag->reg = sendreq->req_rdma[i].btl_reg;
break;
}
}
/* RDMA writes may proceed in parallel to send and to each other, so
* create clone of the convertor for each RDMA fragment
*/
size = hdr->hdr_rdma_offset;
- Split the datatype engine into two parts: an MPI specific part in OMPI and a language agnostic part in OPAL. The convertor is completely moved into OPAL. This offers several benefits as described in RFC http://www.open-mpi.org/community/lists/devel/2009/07/6387.php namely: - Fewer basic types (int* and float* types, boolean and wchar - Fixing naming scheme to ompi-nomenclature. - Usability outside of the ompi-layer. - Due to the fixed nature of simple opal types, their information is completely known at compile time and therefore constified - With fewer datatypes (22), the actual sizes of bit-field types may be reduced from 64 to 32 bits, allowing reorganizing the opal_datatype structure, eliminating holes and keeping data required in convertor (upon send/recv) in one cacheline... This has implications to the convertor-datastructure and other parts of the code. - Several performance tests have been run, the netpipe latency does not change with this patch on Linux/x86-64 on the smoky cluster. - Extensive tests have been done to verify correctness (no new regressions) using: 1. mpi_test_suite on linux/x86-64 using clean ompi-trunk and ompi-ddt: a. running both trunk and ompi-ddt resulted in no differences (except for MPI_SHORT_INT and MPI_TYPE_MIX_LB_UB do now run correctly). b. with --enable-memchecker and running under valgrind (one buglet when run with static found in test-suite, commited) 2. ibm testsuite on linux/x86-64 using clean ompi-trunk and ompi-ddt: all passed (except for the dynamic/ tests failed!! as trunk/MTT) 3. compilation and usage of HDF5 tests on Jaguar using PGI and PathScale compilers. 4. compilation and usage on Scicortex. - Please note, that for the heterogeneous case, (-m32 compiled binaries/ompi), neither ompi-trunk, nor ompi-ddt branch would successfully launch. This commit was SVN r21641.
2009-07-13 08:56:31 +04:00
opal_convertor_clone_with_position(&sendreq->req_send.req_base.req_convertor,
&frag->convertor, 0, &size);
mca_pml_ob1_send_request_put_frag(frag);
}