2015-02-19 23:41:41 +03:00
|
|
|
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
|
2006-07-04 05:20:20 +04:00
|
|
|
/*
|
|
|
|
* Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
|
|
|
|
* University Research and Technology
|
|
|
|
* Corporation. All rights reserved.
|
2013-07-04 12:34:37 +04:00
|
|
|
* Copyright (c) 2004-2013 The University of Tennessee and The University
|
2006-07-04 05:20:20 +04:00
|
|
|
* of Tennessee Research Foundation. All rights
|
|
|
|
* reserved.
|
2015-06-24 06:59:57 +03:00
|
|
|
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
|
2006-07-04 05:20:20 +04:00
|
|
|
* University of Stuttgart. All rights reserved.
|
|
|
|
* Copyright (c) 2004-2006 The Regents of the University of California.
|
|
|
|
* All rights reserved.
|
2012-03-23 02:55:59 +04:00
|
|
|
* Copyright (c) 2012 Sandia National Laboratories. All rights reserved.
|
2015-02-19 23:41:41 +03:00
|
|
|
* Copyright (c) 2015 Los Alamos National Security, LLC. All rights
|
|
|
|
* reserved.
|
2006-07-04 05:20:20 +04:00
|
|
|
* $COPYRIGHT$
|
|
|
|
*
|
|
|
|
* Additional copyrights may follow
|
|
|
|
*
|
|
|
|
* $HEADER$
|
|
|
|
*/
|
|
|
|
|
|
|
|
#ifndef PML_CM_RECVREQ_H
|
|
|
|
#define PML_CM_RECVREQ_H
|
|
|
|
|
2006-07-14 23:32:26 +04:00
|
|
|
#include "pml_cm_request.h"
|
2006-07-04 05:20:20 +04:00
|
|
|
#include "ompi/mca/pml/base/pml_base_recvreq.h"
|
|
|
|
#include "ompi/mca/mtl/mtl.h"
|
|
|
|
|
2006-07-14 23:32:26 +04:00
|
|
|
struct mca_pml_cm_thin_recv_request_t {
|
|
|
|
mca_pml_cm_request_t req_base;
|
2006-12-15 20:54:14 +03:00
|
|
|
mca_mtl_request_t req_mtl; /**< the mtl specific memory. This field should be the last in the struct */
|
2006-07-04 05:20:20 +04:00
|
|
|
};
|
2006-07-14 23:32:26 +04:00
|
|
|
typedef struct mca_pml_cm_thin_recv_request_t mca_pml_cm_thin_recv_request_t;
|
2008-01-09 23:32:39 +03:00
|
|
|
OBJ_CLASS_DECLARATION(mca_pml_cm_thin_recv_request_t);
|
2006-07-04 05:20:20 +04:00
|
|
|
|
2006-07-14 23:32:26 +04:00
|
|
|
struct mca_pml_cm_hvy_recv_request_t {
|
|
|
|
mca_pml_cm_request_t req_base;
|
|
|
|
void *req_addr; /**< pointer to application buffer */
|
|
|
|
size_t req_count; /**< count of user datatype elements */
|
|
|
|
int32_t req_peer; /**< peer process - rank w/in this communicator */
|
|
|
|
int32_t req_tag; /**< user defined tag */
|
2006-11-08 07:58:23 +03:00
|
|
|
void *req_buff; /**< pointer to send buffer - may not be application buffer */
|
|
|
|
size_t req_bytes_packed; /**< packed size of a message given the datatype and count */
|
2006-07-14 23:32:26 +04:00
|
|
|
bool req_blocking;
|
2006-12-15 20:54:14 +03:00
|
|
|
mca_mtl_request_t req_mtl; /**< the mtl specific memory. This field should be the last in the struct */
|
2015-06-24 06:59:57 +03:00
|
|
|
};
|
2006-07-14 23:32:26 +04:00
|
|
|
typedef struct mca_pml_cm_hvy_recv_request_t mca_pml_cm_hvy_recv_request_t;
|
|
|
|
|
2008-01-09 23:32:39 +03:00
|
|
|
OBJ_CLASS_DECLARATION(mca_pml_cm_hvy_recv_request_t);
|
2006-07-04 05:20:20 +04:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Allocate a recv request from the modules free list.
|
|
|
|
*
|
|
|
|
* @param rc (OUT) OMPI_SUCCESS or error status on failure.
|
|
|
|
* @return Receive request.
|
|
|
|
*/
|
2013-07-04 12:34:37 +04:00
|
|
|
#define MCA_PML_CM_THIN_RECV_REQUEST_ALLOC(recvreq) \
|
2015-02-19 23:41:41 +03:00
|
|
|
do { \
|
|
|
|
recvreq = (mca_pml_cm_thin_recv_request_t*) \
|
|
|
|
opal_free_list_get (&mca_pml_base_recv_requests); \
|
2007-07-11 19:10:01 +04:00
|
|
|
recvreq->req_base.req_pml_type = MCA_PML_CM_REQUEST_RECV_THIN; \
|
|
|
|
recvreq->req_mtl.ompi_req = (ompi_request_t*) recvreq; \
|
|
|
|
recvreq->req_mtl.completion_callback = mca_pml_cm_recv_request_completion; \
|
2006-07-14 23:32:26 +04:00
|
|
|
} while (0)
|
|
|
|
|
2013-07-04 12:34:37 +04:00
|
|
|
#define MCA_PML_CM_HVY_RECV_REQUEST_ALLOC(recvreq) \
|
2007-07-11 19:10:01 +04:00
|
|
|
do { \
|
2015-02-19 23:41:41 +03:00
|
|
|
recvreq = (mca_pml_cm_hvy_recv_request_t*) \
|
|
|
|
opal_free_list_get (&mca_pml_base_recv_requests); \
|
2007-07-11 19:10:01 +04:00
|
|
|
recvreq->req_base.req_pml_type = MCA_PML_CM_REQUEST_RECV_HEAVY; \
|
|
|
|
recvreq->req_mtl.ompi_req = (ompi_request_t*) recvreq; \
|
|
|
|
recvreq->req_mtl.completion_callback = mca_pml_cm_recv_request_completion; \
|
2006-07-04 05:20:20 +04:00
|
|
|
} while (0)
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Initialize a receive request with call parameters.
|
|
|
|
*
|
|
|
|
* @param request (IN) Receive request.
|
|
|
|
* @param addr (IN) User buffer.
|
|
|
|
* @param count (IN) Number of elements of indicated datatype.
|
|
|
|
* @param datatype (IN) User defined datatype.
|
|
|
|
* @param src (IN) Source rank w/in the communicator.
|
|
|
|
* @param comm (IN) Communicator.
|
|
|
|
* @param persistent (IN) Is this a ersistent request.
|
|
|
|
*/
|
2015-05-15 02:11:14 +03:00
|
|
|
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
|
2006-07-14 23:32:26 +04:00
|
|
|
#define MCA_PML_CM_THIN_RECV_REQUEST_INIT( request, \
|
|
|
|
ompi_proc, \
|
|
|
|
comm, \
|
|
|
|
src, \
|
|
|
|
datatype, \
|
|
|
|
addr, \
|
|
|
|
count ) \
|
2006-07-04 05:20:20 +04:00
|
|
|
do { \
|
2006-07-14 23:32:26 +04:00
|
|
|
OMPI_REQUEST_INIT(&(request)->req_base.req_ompi, false); \
|
2006-10-27 16:32:36 +04:00
|
|
|
(request)->req_base.req_ompi.req_mpi_object.comm = comm; \
|
2006-07-14 23:32:26 +04:00
|
|
|
(request)->req_base.req_pml_complete = false; \
|
|
|
|
(request)->req_base.req_free_called = false; \
|
2006-11-08 07:58:23 +03:00
|
|
|
request->req_base.req_comm = comm; \
|
|
|
|
request->req_base.req_datatype = datatype; \
|
2006-08-17 20:30:03 +04:00
|
|
|
OBJ_RETAIN(comm); \
|
|
|
|
OBJ_RETAIN(datatype); \
|
2006-07-14 23:32:26 +04:00
|
|
|
\
|
|
|
|
if( MPI_ANY_SOURCE == src ) { \
|
2006-08-17 02:35:53 +04:00
|
|
|
ompi_proc = ompi_proc_local_proc; \
|
2006-07-04 05:20:20 +04:00
|
|
|
} else { \
|
2006-09-21 02:14:46 +04:00
|
|
|
ompi_proc = ompi_comm_peer_lookup( comm, src ); \
|
2006-07-04 05:20:20 +04:00
|
|
|
} \
|
- 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( \
|
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-)
WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL
All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic.
This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
|
|
|
ompi_proc->super.proc_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
|
|
|
&(datatype->super), \
|
2006-07-14 23:32:26 +04:00
|
|
|
count, \
|
|
|
|
addr, \
|
|
|
|
0, \
|
|
|
|
&(request)->req_base.req_convertor ); \
|
2006-07-04 05:20:20 +04:00
|
|
|
} while(0)
|
2015-05-15 02:11:14 +03:00
|
|
|
#else
|
|
|
|
#define MCA_PML_CM_THIN_RECV_REQUEST_INIT( request, \
|
|
|
|
ompi_proc, \
|
|
|
|
comm, \
|
|
|
|
src, \
|
|
|
|
datatype, \
|
|
|
|
addr, \
|
|
|
|
count ) \
|
|
|
|
do { \
|
|
|
|
OMPI_REQUEST_INIT(&(request)->req_base.req_ompi, false); \
|
|
|
|
(request)->req_base.req_ompi.req_mpi_object.comm = comm; \
|
|
|
|
(request)->req_base.req_pml_complete = false; \
|
|
|
|
(request)->req_base.req_free_called = false; \
|
|
|
|
request->req_base.req_comm = comm; \
|
|
|
|
request->req_base.req_datatype = datatype; \
|
|
|
|
OBJ_RETAIN(comm); \
|
|
|
|
OBJ_RETAIN(datatype); \
|
|
|
|
\
|
|
|
|
opal_convertor_copy_and_prepare_for_recv( \
|
|
|
|
ompi_mpi_local_convertor, \
|
|
|
|
&(datatype->super), \
|
|
|
|
count, \
|
|
|
|
addr, \
|
|
|
|
0, \
|
|
|
|
&(request)->req_base.req_convertor ); \
|
|
|
|
} while(0)
|
|
|
|
#endif
|
2006-07-04 05:20:20 +04:00
|
|
|
|
2015-05-15 02:11:14 +03:00
|
|
|
#if OPAL_ENABLE_HETEROGENEOUS_SUPPORT
|
2006-07-14 23:32:26 +04:00
|
|
|
#define MCA_PML_CM_HVY_RECV_REQUEST_INIT( request, \
|
|
|
|
ompi_proc, \
|
|
|
|
comm, \
|
|
|
|
tag, \
|
|
|
|
src, \
|
|
|
|
datatype, \
|
|
|
|
addr, \
|
|
|
|
count, \
|
|
|
|
persistent) \
|
|
|
|
do { \
|
|
|
|
OMPI_REQUEST_INIT(&(request)->req_base.req_ompi, persistent); \
|
2006-10-27 16:32:36 +04:00
|
|
|
(request)->req_base.req_ompi.req_mpi_object.comm = comm; \
|
2006-08-28 22:59:16 +04:00
|
|
|
(request)->req_base.req_pml_complete = OPAL_INT_TO_BOOL(persistent); \
|
2006-07-14 23:32:26 +04:00
|
|
|
(request)->req_base.req_free_called = false; \
|
2006-11-08 07:58:23 +03:00
|
|
|
request->req_base.req_comm = comm; \
|
|
|
|
request->req_base.req_datatype = datatype; \
|
2006-07-14 23:32:26 +04:00
|
|
|
request->req_tag = tag; \
|
|
|
|
request->req_peer = src; \
|
|
|
|
request->req_addr = addr; \
|
|
|
|
request->req_count = count; \
|
2006-08-17 20:30:03 +04:00
|
|
|
OBJ_RETAIN(comm); \
|
|
|
|
OBJ_RETAIN(datatype); \
|
2006-07-14 23:32:26 +04:00
|
|
|
\
|
|
|
|
if( MPI_ANY_SOURCE == src ) { \
|
2006-08-17 02:35:53 +04:00
|
|
|
ompi_proc = ompi_proc_local_proc; \
|
2006-07-14 23:32:26 +04:00
|
|
|
} else { \
|
2006-09-21 02:14:46 +04:00
|
|
|
ompi_proc = ompi_comm_peer_lookup( comm, src ); \
|
2006-07-14 23:32:26 +04:00
|
|
|
} \
|
- 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( \
|
George did the work and deserves all the credit for it. Ralph did the merge, and deserves whatever blame results from errors in it :-)
WHAT: Open our low-level communication infrastructure by moving all necessary components (btl/rcache/allocator/mpool) down in OPAL
All the components required for inter-process communications are currently deeply integrated in the OMPI layer. Several groups/institutions have express interest in having a more generic communication infrastructure, without all the OMPI layer dependencies. This communication layer should be made available at a different software level, available to all layers in the Open MPI software stack. As an example, our ORTE layer could replace the current OOB and instead use the BTL directly, gaining access to more reactive network interfaces than TCP. Similarly, external software libraries could take advantage of our highly optimized AM (active message) communication layer for their own purpose. UTK with support from Sandia, developped a version of Open MPI where the entire communication infrastucture has been moved down to OPAL (btl/rcache/allocator/mpool). Most of the moved components have been updated to match the new schema, with few exceptions (mainly BTLs where I have no way of compiling/testing them). Thus, the completion of this RFC is tied to being able to completing this move for all BTLs. For this we need help from the rest of the Open MPI community, especially those supporting some of the BTLs. A non-exhaustive list of BTLs that qualify here is: mx, portals4, scif, udapl, ugni, usnic.
This commit was SVN r32317.
2014-07-26 04:47:28 +04:00
|
|
|
ompi_proc->super.proc_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
|
|
|
&(datatype->super), \
|
2006-07-14 23:32:26 +04:00
|
|
|
count, \
|
|
|
|
addr, \
|
|
|
|
0, \
|
|
|
|
&(request)->req_base.req_convertor ); \
|
|
|
|
} while(0)
|
2015-05-15 02:11:14 +03:00
|
|
|
#else
|
|
|
|
#define MCA_PML_CM_HVY_RECV_REQUEST_INIT( request, \
|
|
|
|
ompi_proc, \
|
|
|
|
comm, \
|
|
|
|
tag, \
|
|
|
|
src, \
|
|
|
|
datatype, \
|
|
|
|
addr, \
|
|
|
|
count, \
|
|
|
|
persistent) \
|
|
|
|
do { \
|
|
|
|
OMPI_REQUEST_INIT(&(request)->req_base.req_ompi, persistent); \
|
|
|
|
(request)->req_base.req_ompi.req_mpi_object.comm = comm; \
|
|
|
|
(request)->req_base.req_pml_complete = OPAL_INT_TO_BOOL(persistent); \
|
|
|
|
(request)->req_base.req_free_called = false; \
|
|
|
|
request->req_base.req_comm = comm; \
|
|
|
|
request->req_base.req_datatype = datatype; \
|
|
|
|
request->req_tag = tag; \
|
|
|
|
request->req_peer = src; \
|
|
|
|
request->req_addr = addr; \
|
|
|
|
request->req_count = count; \
|
|
|
|
OBJ_RETAIN(comm); \
|
|
|
|
OBJ_RETAIN(datatype); \
|
|
|
|
\
|
|
|
|
opal_convertor_copy_and_prepare_for_recv( \
|
|
|
|
ompi_mpi_local_convertor, \
|
|
|
|
&(datatype->super), \
|
|
|
|
count, \
|
|
|
|
addr, \
|
|
|
|
0, \
|
|
|
|
&(request)->req_base.req_convertor ); \
|
|
|
|
} while(0)
|
|
|
|
#endif
|
2006-07-04 05:20:20 +04:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Start an initialized request.
|
|
|
|
*
|
|
|
|
* @param request Receive request.
|
|
|
|
* @return OMPI_SUCESS or error status on failure.
|
|
|
|
*/
|
2006-07-14 23:32:26 +04:00
|
|
|
#define MCA_PML_CM_THIN_RECV_REQUEST_START(request, comm, tag, src, ret) \
|
2006-07-04 05:20:20 +04:00
|
|
|
do { \
|
|
|
|
/* init/re-init the request */ \
|
2006-07-14 23:32:26 +04:00
|
|
|
request->req_base.req_pml_complete = false; \
|
|
|
|
request->req_base.req_ompi.req_complete = false; \
|
|
|
|
request->req_base.req_ompi.req_state = OMPI_REQUEST_ACTIVE; \
|
|
|
|
\
|
|
|
|
/* always set the req_status.MPI_TAG to ANY_TAG before starting the \
|
|
|
|
* request. This field is used if cancelled to find out if the request \
|
|
|
|
* has been matched or not. \
|
|
|
|
*/ \
|
|
|
|
request->req_base.req_ompi.req_status.MPI_TAG = OMPI_ANY_TAG; \
|
|
|
|
request->req_base.req_ompi.req_status.MPI_ERROR = OMPI_SUCCESS; \
|
|
|
|
request->req_base.req_ompi.req_status._cancelled = 0; \
|
|
|
|
ret = OMPI_MTL_CALL(irecv(ompi_mtl, \
|
|
|
|
comm, \
|
|
|
|
src, \
|
|
|
|
tag, \
|
|
|
|
&recvreq->req_base.req_convertor, \
|
2006-12-15 20:54:14 +03:00
|
|
|
&recvreq->req_mtl)); \
|
2006-07-14 23:32:26 +04:00
|
|
|
} while (0)
|
|
|
|
|
2012-03-23 02:55:59 +04:00
|
|
|
#define MCA_PML_CM_THIN_RECV_REQUEST_MATCHED_START(request, message, ret) \
|
|
|
|
do { \
|
|
|
|
/* init/re-init the request */ \
|
|
|
|
request->req_base.req_pml_complete = false; \
|
|
|
|
request->req_base.req_ompi.req_complete = false; \
|
|
|
|
request->req_base.req_ompi.req_state = OMPI_REQUEST_ACTIVE; \
|
|
|
|
\
|
|
|
|
/* always set the req_status.MPI_TAG to ANY_TAG before starting the \
|
|
|
|
* request. This field is used if cancelled to find out if the request \
|
|
|
|
* has been matched or not. \
|
|
|
|
*/ \
|
|
|
|
request->req_base.req_ompi.req_status.MPI_TAG = OMPI_ANY_TAG; \
|
|
|
|
request->req_base.req_ompi.req_status.MPI_ERROR = OMPI_SUCCESS; \
|
|
|
|
request->req_base.req_ompi.req_status._cancelled = 0; \
|
|
|
|
ret = OMPI_MTL_CALL(imrecv(ompi_mtl, \
|
|
|
|
&recvreq->req_base.req_convertor, \
|
|
|
|
message, \
|
|
|
|
&recvreq->req_mtl)); \
|
|
|
|
} while (0)
|
|
|
|
|
2006-07-14 23:32:26 +04:00
|
|
|
|
|
|
|
#define MCA_PML_CM_HVY_RECV_REQUEST_START(request, ret) \
|
|
|
|
do { \
|
2008-06-09 18:53:58 +04:00
|
|
|
/* opal_output(0, "posting hvy request %d\n", request); */ \
|
2006-07-14 23:32:26 +04:00
|
|
|
/* init/re-init the request */ \
|
|
|
|
request->req_base.req_pml_complete = false; \
|
|
|
|
request->req_base.req_ompi.req_complete = false; \
|
|
|
|
request->req_base.req_ompi.req_state = OMPI_REQUEST_ACTIVE; \
|
|
|
|
\
|
|
|
|
/* always set the req_status.MPI_TAG to ANY_TAG before starting the \
|
|
|
|
* request. This field is used if cancelled to find out if the request \
|
|
|
|
* has been matched or not. \
|
|
|
|
*/ \
|
|
|
|
request->req_base.req_ompi.req_status.MPI_TAG = OMPI_ANY_TAG; \
|
|
|
|
request->req_base.req_ompi.req_status.MPI_ERROR = OMPI_SUCCESS; \
|
|
|
|
request->req_base.req_ompi.req_status._cancelled = 0; \
|
2006-07-04 05:20:20 +04:00
|
|
|
ret = OMPI_MTL_CALL(irecv(ompi_mtl, \
|
2006-11-08 07:58:23 +03:00
|
|
|
request->req_base.req_comm, \
|
2006-07-14 23:32:26 +04:00
|
|
|
request->req_peer, \
|
|
|
|
request->req_tag, \
|
|
|
|
&recvreq->req_base.req_convertor, \
|
2006-12-15 20:54:14 +03:00
|
|
|
&recvreq->req_mtl)); \
|
2006-07-04 05:20:20 +04:00
|
|
|
} while (0)
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Mark the request as completed at MPI level for internal purposes.
|
|
|
|
*
|
|
|
|
* @param recvreq (IN) Receive request.
|
|
|
|
*/
|
2006-07-14 23:32:26 +04:00
|
|
|
#define MCA_PML_CM_THIN_RECV_REQUEST_MPI_COMPLETE( recvreq ) \
|
2006-07-04 05:20:20 +04:00
|
|
|
do { \
|
2008-05-30 07:58:39 +04:00
|
|
|
ompi_request_complete( &(recvreq->req_base.req_ompi), true ); \
|
2006-07-04 05:20:20 +04:00
|
|
|
} while (0)
|
2015-06-24 06:59:57 +03:00
|
|
|
|
2006-07-14 23:32:26 +04:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Return a recv request to the modules free list.
|
|
|
|
*
|
|
|
|
* @param recvreq (IN) Receive request.
|
|
|
|
*/
|
|
|
|
#define MCA_PML_CM_THIN_RECV_REQUEST_PML_COMPLETE(recvreq) \
|
|
|
|
do { \
|
|
|
|
assert( false == recvreq->req_base.req_pml_complete ); \
|
|
|
|
\
|
|
|
|
OPAL_THREAD_LOCK(&ompi_request_lock); \
|
|
|
|
\
|
|
|
|
if( true == recvreq->req_base.req_free_called ) { \
|
|
|
|
MCA_PML_CM_THIN_RECV_REQUEST_RETURN( recvreq ); \
|
|
|
|
} else { \
|
|
|
|
recvreq->req_base.req_pml_complete = true; \
|
2008-05-30 07:58:39 +04:00
|
|
|
ompi_request_complete( &(recvreq->req_base.req_ompi), true ); \
|
2006-07-14 23:32:26 +04:00
|
|
|
} \
|
|
|
|
OPAL_THREAD_UNLOCK(&ompi_request_lock); \
|
|
|
|
} while(0)
|
|
|
|
|
|
|
|
|
2006-07-04 05:20:20 +04:00
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Return a recv request to the modules free list.
|
|
|
|
*
|
|
|
|
* @param recvreq (IN) Receive request.
|
|
|
|
*/
|
2006-07-14 23:32:26 +04:00
|
|
|
#define MCA_PML_CM_HVY_RECV_REQUEST_PML_COMPLETE(recvreq) \
|
2006-07-04 05:20:20 +04:00
|
|
|
do { \
|
2006-07-14 23:32:26 +04:00
|
|
|
assert( false == recvreq->req_base.req_pml_complete ); \
|
2006-07-04 05:20:20 +04:00
|
|
|
\
|
|
|
|
OPAL_THREAD_LOCK(&ompi_request_lock); \
|
|
|
|
\
|
2006-07-14 23:32:26 +04:00
|
|
|
if( true == recvreq->req_base.req_free_called ) { \
|
|
|
|
MCA_PML_CM_HVY_RECV_REQUEST_RETURN( recvreq ); \
|
2006-07-04 05:20:20 +04:00
|
|
|
} else { \
|
|
|
|
/* initialize request status */ \
|
2006-07-14 23:32:26 +04:00
|
|
|
if(recvreq->req_base.req_ompi.req_persistent) { \
|
2006-07-06 19:13:26 +04:00
|
|
|
/* rewind convertor */ \
|
|
|
|
size_t offset = 0; \
|
- 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(&recvreq->req_base.req_convertor, &offset); \
|
2006-07-06 19:13:26 +04:00
|
|
|
} \
|
2006-07-14 23:32:26 +04:00
|
|
|
recvreq->req_base.req_pml_complete = true; \
|
2008-05-30 07:58:39 +04:00
|
|
|
ompi_request_complete( &(recvreq->req_base.req_ompi), true ); \
|
2006-07-04 05:20:20 +04:00
|
|
|
} \
|
|
|
|
OPAL_THREAD_UNLOCK(&ompi_request_lock); \
|
|
|
|
} while(0)
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Free the PML receive request
|
|
|
|
*/
|
2006-08-17 20:30:03 +04:00
|
|
|
#define MCA_PML_CM_HVY_RECV_REQUEST_RETURN(recvreq) \
|
|
|
|
{ \
|
2006-11-08 07:58:23 +03:00
|
|
|
OBJ_RELEASE((recvreq)->req_base.req_comm); \
|
|
|
|
OBJ_RELEASE((recvreq)->req_base.req_datatype); \
|
2006-07-14 23:32:26 +04:00
|
|
|
OMPI_REQUEST_FINI(&(recvreq)->req_base.req_ompi); \
|
- 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( &((recvreq)->req_base.req_convertor) ); \
|
2015-02-19 23:41:41 +03:00
|
|
|
opal_free_list_return ( &mca_pml_base_recv_requests, \
|
|
|
|
(opal_free_list_item_t*)(recvreq)); \
|
2006-07-04 05:20:20 +04:00
|
|
|
}
|
|
|
|
|
2006-07-14 23:32:26 +04:00
|
|
|
/**
|
|
|
|
* Free the PML receive request
|
|
|
|
*/
|
2006-08-17 20:30:03 +04:00
|
|
|
#define MCA_PML_CM_THIN_RECV_REQUEST_RETURN(recvreq) \
|
|
|
|
{ \
|
2006-11-08 07:58:23 +03:00
|
|
|
OBJ_RELEASE((recvreq)->req_base.req_comm); \
|
|
|
|
OBJ_RELEASE((recvreq)->req_base.req_datatype); \
|
2006-07-14 23:32:26 +04:00
|
|
|
OMPI_REQUEST_FINI(&(recvreq)->req_base.req_ompi); \
|
- 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( &((recvreq)->req_base.req_convertor) ); \
|
2015-02-19 23:41:41 +03:00
|
|
|
opal_free_list_return ( &mca_pml_base_recv_requests, \
|
|
|
|
(opal_free_list_item_t*)(recvreq)); \
|
2006-07-14 23:32:26 +04:00
|
|
|
}
|
|
|
|
|
2007-07-11 19:10:01 +04:00
|
|
|
extern void mca_pml_cm_recv_request_completion(struct mca_mtl_request_t *mtl_request);
|
2006-07-04 05:20:20 +04:00
|
|
|
|
|
|
|
#endif
|
2006-08-16 01:12:03 +04:00
|
|
|
|
|
|
|
|