2006-07-18 02:08:55 +04:00
|
|
|
/*
|
|
|
|
* Copyright (c) 2004-2005 The Trustees of Indiana University.
|
|
|
|
* All rights reserved.
|
|
|
|
* Copyright (c) 2004-2005 The Trustees of the University of Tennessee.
|
|
|
|
* 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.
|
2007-05-24 19:41:24 +04:00
|
|
|
* Copyright (c) 2007 Los Alamos National Security, LLC. All rights
|
|
|
|
* reserved.
|
2010-07-20 22:45:48 +04:00
|
|
|
* Copyright (c) 2010 Cisco Systems, Inc. All rights reserved.
|
2006-07-18 02:08:55 +04:00
|
|
|
* $COPYRIGHT$
|
|
|
|
*
|
|
|
|
* Additional copyrights may follow
|
|
|
|
*
|
|
|
|
* $HEADER$
|
|
|
|
*/
|
|
|
|
|
2006-08-03 04:10:19 +04:00
|
|
|
#ifndef OMPI_OSC_RDMA_SENDREQ_H
|
|
|
|
#define OMPI_OSC_RDMA_SENDREQ_H
|
2006-07-18 02:08:55 +04:00
|
|
|
|
|
|
|
#include "osc_rdma.h"
|
|
|
|
#include "osc_rdma_longreq.h"
|
|
|
|
|
|
|
|
#include "opal/class/opal_list.h"
|
- 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
|
|
|
#include "ompi/datatype/ompi_datatype.h"
|
|
|
|
#include "opal/datatype/opal_convertor.h"
|
2006-07-18 02:08:55 +04:00
|
|
|
#include "ompi/communicator/communicator.h"
|
|
|
|
#include "ompi/proc/proc.h"
|
2008-04-16 17:29:55 +04:00
|
|
|
#include "ompi/memchecker.h"
|
2006-07-18 02:08:55 +04:00
|
|
|
|
|
|
|
typedef enum {
|
2006-08-03 04:10:19 +04:00
|
|
|
OMPI_OSC_RDMA_GET,
|
|
|
|
OMPI_OSC_RDMA_ACC,
|
|
|
|
OMPI_OSC_RDMA_PUT
|
2006-07-18 02:08:55 +04:00
|
|
|
} ompi_osc_rdma_req_type_t;
|
|
|
|
|
|
|
|
|
|
|
|
struct ompi_osc_rdma_sendreq_t {
|
|
|
|
ompi_request_t super;
|
|
|
|
|
2007-07-06 01:40:06 +04:00
|
|
|
int req_refcount;
|
|
|
|
|
2006-07-18 02:08:55 +04:00
|
|
|
/** type of sendreq (from ompi_osc_rdma_req_type_t) */
|
|
|
|
ompi_osc_rdma_req_type_t req_type;
|
|
|
|
/** pointer to the module that created the sendreq */
|
|
|
|
ompi_osc_rdma_module_t *req_module;
|
|
|
|
|
|
|
|
/** Datatype for the origin side of the operation */
|
|
|
|
struct ompi_datatype_t *req_origin_datatype;
|
|
|
|
/** Convertor for the origin side of the operation. Setup for
|
|
|
|
either send (Put / Accumulate) or receive (Get) */
|
- 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_t req_origin_convertor;
|
2006-07-18 02:08:55 +04:00
|
|
|
/** packed size of message on the origin side */
|
|
|
|
size_t req_origin_bytes_packed;
|
|
|
|
|
|
|
|
/** rank in module's communicator for target of operation */
|
|
|
|
int req_target_rank;
|
|
|
|
/** pointer to the proc structure for the target of the operation */
|
|
|
|
ompi_proc_t *req_target_proc;
|
|
|
|
|
|
|
|
/** displacement on target */
|
2010-07-20 22:45:48 +04:00
|
|
|
OPAL_PTRDIFF_TYPE req_target_disp;
|
2006-07-18 02:08:55 +04:00
|
|
|
/** datatype count on target */
|
|
|
|
int req_target_count;
|
|
|
|
/** datatype on target */
|
|
|
|
struct ompi_datatype_t *req_target_datatype;
|
|
|
|
|
|
|
|
/** op index on the target */
|
|
|
|
int req_op_id;
|
2007-07-05 07:32:32 +04:00
|
|
|
|
|
|
|
mca_btl_base_segment_t remote_segs[1];
|
2006-07-18 02:08:55 +04:00
|
|
|
};
|
|
|
|
typedef struct ompi_osc_rdma_sendreq_t ompi_osc_rdma_sendreq_t;
|
|
|
|
OBJ_CLASS_DECLARATION(ompi_osc_rdma_sendreq_t);
|
|
|
|
|
|
|
|
|
|
|
|
/** allocate and populate a sendreq structure. Both datatypes are
|
|
|
|
RETAINed for the life of the sendreq */
|
|
|
|
int
|
|
|
|
ompi_osc_rdma_sendreq_alloc_init(ompi_osc_rdma_req_type_t req_type,
|
|
|
|
void *origin_addr, int origin_count,
|
|
|
|
struct ompi_datatype_t *origin_dt,
|
2010-07-20 22:45:48 +04:00
|
|
|
int target, OPAL_PTRDIFF_TYPE target_disp,
|
|
|
|
int target_count,
|
2006-07-18 02:08:55 +04:00
|
|
|
struct ompi_datatype_t *target_datatype,
|
|
|
|
ompi_osc_rdma_module_t *module,
|
|
|
|
ompi_osc_rdma_sendreq_t **sendreq);
|
|
|
|
|
|
|
|
static inline int
|
|
|
|
ompi_osc_rdma_sendreq_alloc(ompi_osc_rdma_module_t *module,
|
|
|
|
int target_rank,
|
|
|
|
ompi_osc_rdma_sendreq_t **sendreq)
|
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
opal_free_list_item_t *item;
|
2007-05-24 19:41:24 +04:00
|
|
|
ompi_proc_t *proc = ompi_comm_peer_lookup( module->m_comm, target_rank );
|
2006-07-18 02:08:55 +04:00
|
|
|
|
|
|
|
/* BWB - FIX ME - is this really the right return code? */
|
|
|
|
if (NULL == proc) return OMPI_ERR_OUT_OF_RESOURCE;
|
|
|
|
|
2007-05-24 19:41:24 +04:00
|
|
|
OPAL_FREE_LIST_GET(&mca_osc_rdma_component.c_sendreqs,
|
2006-07-18 02:08:55 +04:00
|
|
|
item, ret);
|
|
|
|
if (OMPI_SUCCESS != ret) return ret;
|
|
|
|
*sendreq = (ompi_osc_rdma_sendreq_t*) item;
|
|
|
|
|
|
|
|
(*sendreq)->req_module = module;
|
|
|
|
(*sendreq)->req_target_rank = target_rank;
|
|
|
|
(*sendreq)->req_target_proc = proc;
|
2007-07-06 01:40:06 +04:00
|
|
|
(*sendreq)->req_refcount = 1;
|
2006-07-18 02:08:55 +04:00
|
|
|
|
|
|
|
return OMPI_SUCCESS;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static inline int
|
|
|
|
ompi_osc_rdma_sendreq_init_origin(ompi_osc_rdma_sendreq_t *sendreq,
|
|
|
|
ompi_osc_rdma_req_type_t req_type,
|
|
|
|
void *origin_addr,
|
|
|
|
int origin_count,
|
|
|
|
struct ompi_datatype_t *origin_dt)
|
|
|
|
{
|
|
|
|
OBJ_RETAIN(origin_dt);
|
|
|
|
sendreq->req_origin_datatype = origin_dt;
|
|
|
|
sendreq->req_type = req_type;
|
|
|
|
|
2006-08-03 04:10:19 +04:00
|
|
|
if (req_type != OMPI_OSC_RDMA_GET) {
|
- 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_copy_and_prepare_for_send(sendreq->req_target_proc->proc_convertor,
|
|
|
|
&(origin_dt->super),
|
2006-07-18 02:08:55 +04:00
|
|
|
origin_count,
|
|
|
|
origin_addr,
|
|
|
|
0,
|
|
|
|
&(sendreq->req_origin_convertor));
|
- 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_get_packed_size(&sendreq->req_origin_convertor,
|
2006-07-18 02:08:55 +04:00
|
|
|
&sendreq->req_origin_bytes_packed);
|
|
|
|
} else {
|
- 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_copy_and_prepare_for_recv(sendreq->req_target_proc->proc_convertor,
|
|
|
|
&(origin_dt->super),
|
2006-07-18 02:08:55 +04:00
|
|
|
origin_count,
|
|
|
|
origin_addr,
|
|
|
|
0,
|
|
|
|
&(sendreq->req_origin_convertor));
|
- 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_get_packed_size(&sendreq->req_origin_convertor,
|
2006-07-18 02:08:55 +04:00
|
|
|
&sendreq->req_origin_bytes_packed);
|
|
|
|
}
|
|
|
|
|
|
|
|
return OMPI_SUCCESS;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static inline int
|
|
|
|
ompi_osc_rdma_sendreq_init_target(ompi_osc_rdma_sendreq_t *sendreq,
|
2010-07-20 22:45:48 +04:00
|
|
|
OPAL_PTRDIFF_TYPE target_disp,
|
2006-07-18 02:08:55 +04:00
|
|
|
int target_count,
|
|
|
|
struct ompi_datatype_t *target_datatype)
|
|
|
|
{
|
|
|
|
OBJ_RETAIN(target_datatype);
|
|
|
|
|
|
|
|
sendreq->req_target_disp = target_disp;
|
|
|
|
sendreq->req_target_count = target_count;
|
|
|
|
sendreq->req_target_datatype = target_datatype;
|
|
|
|
|
|
|
|
return OMPI_SUCCESS;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static inline int
|
|
|
|
ompi_osc_rdma_sendreq_free(ompi_osc_rdma_sendreq_t *sendreq)
|
|
|
|
{
|
2007-07-06 01:40:06 +04:00
|
|
|
if (0 == (--sendreq->req_refcount)) {
|
2008-04-16 17:29:55 +04:00
|
|
|
MEMCHECKER(
|
|
|
|
memchecker_convertor_call(&opal_memchecker_base_mem_defined,
|
|
|
|
&sendreq->req_origin_convertor);
|
|
|
|
);
|
- 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_cleanup(&sendreq->req_origin_convertor);
|
2007-07-06 01:40:06 +04:00
|
|
|
|
|
|
|
OBJ_RELEASE(sendreq->req_target_datatype);
|
|
|
|
OBJ_RELEASE(sendreq->req_origin_datatype);
|
2006-07-18 02:08:55 +04:00
|
|
|
|
2007-07-06 01:40:06 +04:00
|
|
|
OPAL_FREE_LIST_RETURN(&mca_osc_rdma_component.c_sendreqs,
|
|
|
|
(opal_list_item_t*) sendreq);
|
|
|
|
}
|
2006-07-18 02:08:55 +04:00
|
|
|
|
|
|
|
return OMPI_SUCCESS;
|
|
|
|
}
|
|
|
|
|
2006-08-03 04:10:19 +04:00
|
|
|
#endif /* OMPI_OSC_RDMA_SENDREQ_H */
|