2008-04-02 23:02:33 +04:00
|
|
|
/*
|
|
|
|
* Copyright (c) 2007-2008 UT-Battelle, LLC
|
|
|
|
* $COPYRIGHT$
|
|
|
|
*
|
|
|
|
* Additional copyrights may follow
|
|
|
|
*
|
|
|
|
* $HEADER$
|
|
|
|
*/
|
|
|
|
/** @file */
|
|
|
|
|
|
|
|
#include "ompi_config.h"
|
|
|
|
|
|
|
|
#include "ompi/constants.h"
|
|
|
|
#include "coll_sm2.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"
|
2008-04-02 23:02:33 +04:00
|
|
|
#include "ompi/communicator/communicator.h"
|
|
|
|
/* debug
|
|
|
|
#include "opal/sys/timer.h"
|
|
|
|
|
|
|
|
extern uint64_t timers[7];
|
|
|
|
end debug */
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Shared memory blocking allreduce.
|
|
|
|
*/
|
|
|
|
|
|
|
|
static
|
|
|
|
int mca_coll_sm2_fanout(void *buf, int count,
|
|
|
|
struct ompi_datatype_t *dtype, int root,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module)
|
2008-04-02 23:02:33 +04:00
|
|
|
{
|
|
|
|
/* local variables */
|
|
|
|
int rc=OMPI_SUCCESS,n_dts_per_buffer,n_data_segments,stripe_number;
|
|
|
|
int comm_size,process_shift,my_node_index;
|
|
|
|
int my_rank;
|
|
|
|
int count_processed,count_this_stripe;
|
|
|
|
int my_fanout_parent;
|
|
|
|
size_t message_extent,dt_extent,ctl_size,len_data_buffer;
|
|
|
|
long long tag;
|
|
|
|
volatile char * my_data_pointer;
|
|
|
|
volatile char * parent_data_pointer;
|
|
|
|
mca_coll_sm2_nb_request_process_shared_mem_t *my_ctl_pointer;
|
|
|
|
volatile mca_coll_sm2_nb_request_process_shared_mem_t * parent_ctl_pointer;
|
|
|
|
mca_coll_sm2_module_t *sm_module;
|
|
|
|
tree_node_t *my_fanout_read_tree;
|
|
|
|
sm_work_buffer_t *sm_buffer_desc;
|
|
|
|
|
|
|
|
sm_module=(mca_coll_sm2_module_t *) module;
|
|
|
|
|
|
|
|
|
|
|
|
/* compute process shift */
|
|
|
|
my_rank=ompi_comm_rank(comm);
|
|
|
|
comm_size=ompi_comm_size(comm);
|
|
|
|
process_shift=root;
|
|
|
|
my_node_index=my_rank-root;
|
|
|
|
/* wrap around */
|
|
|
|
if(0 > my_node_index ) {
|
|
|
|
my_node_index+=comm_size;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/* get size of data needed - same layout as user data, so that
|
|
|
|
* we can apply the reudction routines directly on these buffers
|
|
|
|
*/
|
- 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
|
|
|
rc=ompi_datatype_type_extent(dtype, &dt_extent);
|
2008-04-02 23:02:33 +04:00
|
|
|
if( OMPI_SUCCESS != rc ) {
|
|
|
|
goto Error;
|
|
|
|
}
|
|
|
|
message_extent=dt_extent*count;
|
|
|
|
|
|
|
|
/* lenght of control and data regions */
|
|
|
|
ctl_size=sm_module->ctl_memory_per_proc_per_segment;
|
|
|
|
len_data_buffer=sm_module->data_memory_per_proc_per_segment;
|
|
|
|
|
|
|
|
/* number of data types copies that the scratch buffer can hold */
|
|
|
|
n_dts_per_buffer=((int) len_data_buffer)/dt_extent;
|
|
|
|
if ( 0 == n_dts_per_buffer ) {
|
|
|
|
rc=OMPI_ERROR;
|
|
|
|
goto Error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* compute number of stripes needed to process this collective */
|
|
|
|
n_data_segments=(count+n_dts_per_buffer -1 ) / n_dts_per_buffer ;
|
|
|
|
|
|
|
|
/* get my node for the reduction tree */
|
|
|
|
my_fanout_read_tree=&(sm_module->fanout_read_tree[my_node_index]);
|
|
|
|
my_fanout_parent=my_fanout_read_tree->parent_rank+process_shift;
|
|
|
|
if( comm_size <= my_fanout_parent ){
|
|
|
|
my_fanout_parent-=comm_size;
|
|
|
|
}
|
|
|
|
|
|
|
|
count_processed=0;
|
|
|
|
|
|
|
|
/* get a pointer to the shared-memory working buffer */
|
|
|
|
/* NOTE: starting with a rather synchronous approach */
|
|
|
|
for( stripe_number=0 ; stripe_number < n_data_segments ; stripe_number++ ) {
|
|
|
|
|
|
|
|
/* get unique tag for this stripe - assume only one collective
|
|
|
|
* per communicator at a given time, so no locking needed
|
|
|
|
* for atomic update of the tag */
|
|
|
|
tag=sm_module->collective_tag;
|
|
|
|
sm_module->collective_tag++;
|
|
|
|
|
|
|
|
sm_buffer_desc=alloc_sm2_shared_buffer(sm_module);
|
|
|
|
|
|
|
|
/* get number of elements to process in this stripe */
|
|
|
|
count_this_stripe=n_dts_per_buffer;
|
|
|
|
if( count_processed + count_this_stripe > count )
|
|
|
|
count_this_stripe=count-count_processed;
|
|
|
|
|
|
|
|
/* offset to data segment */
|
|
|
|
my_ctl_pointer=sm_buffer_desc->proc_memory[my_rank].control_region;
|
|
|
|
my_data_pointer=sm_buffer_desc->proc_memory[my_rank].data_segment;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Fan out from root - let the memory copies at each
|
|
|
|
* stage help reduce memory contention.
|
|
|
|
*/
|
|
|
|
if( ROOT_NODE == my_fanout_read_tree->my_node_type ) {
|
|
|
|
|
|
|
|
/* copy data to user supplied buffer */
|
- 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
|
|
|
rc=ompi_datatype_copy_content_same_ddt(dtype, count_this_stripe,
|
2008-04-02 23:02:33 +04:00
|
|
|
(char *)my_data_pointer,
|
|
|
|
(char *)((char *)buf+dt_extent*count_processed));
|
|
|
|
if( 0 != rc ) {
|
|
|
|
return OMPI_ERROR;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* I am the root - so copy signal children, and then
|
|
|
|
* start reading
|
|
|
|
*/
|
|
|
|
MB();
|
|
|
|
my_ctl_pointer->flag=tag;
|
|
|
|
|
|
|
|
} else if( LEAF_NODE == my_fanout_read_tree->my_node_type ) {
|
|
|
|
|
|
|
|
parent_data_pointer=
|
|
|
|
sm_buffer_desc->proc_memory[my_fanout_parent].data_segment;
|
|
|
|
parent_ctl_pointer=
|
|
|
|
sm_buffer_desc->proc_memory[my_fanout_parent].control_region;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* wait on Parent to signal that data is ready
|
|
|
|
*/
|
|
|
|
while( parent_ctl_pointer->flag != tag) {
|
|
|
|
opal_progress();
|
|
|
|
}
|
|
|
|
|
|
|
|
/* copy data to user supplied buffer */
|
- 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
|
|
|
rc=ompi_datatype_copy_content_same_ddt(dtype, count_this_stripe,
|
2008-04-02 23:02:33 +04:00
|
|
|
(char *)buf+dt_extent*count_processed,
|
|
|
|
(char *)parent_data_pointer);
|
|
|
|
if( 0 != rc ) {
|
|
|
|
return OMPI_ERROR;
|
|
|
|
}
|
|
|
|
|
|
|
|
} else {
|
|
|
|
/* interior nodes */
|
|
|
|
|
|
|
|
parent_data_pointer=
|
|
|
|
sm_buffer_desc->proc_memory[my_fanout_parent].data_segment;
|
|
|
|
parent_ctl_pointer=
|
|
|
|
sm_buffer_desc->proc_memory[my_fanout_parent].control_region;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* wait on Parent to signal that data is ready
|
|
|
|
*/
|
|
|
|
while( parent_ctl_pointer->flag != tag) {
|
|
|
|
opal_progress();
|
|
|
|
}
|
|
|
|
|
|
|
|
/* copy the data to my shared buffer, for access by children */
|
- 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
|
|
|
rc=ompi_datatype_copy_content_same_ddt(dtype, count_this_stripe,
|
2008-04-02 23:02:33 +04:00
|
|
|
(char *)my_data_pointer,(char *)parent_data_pointer);
|
|
|
|
if( 0 != rc ) {
|
|
|
|
return OMPI_ERROR;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* set memory barriet to make sure data is in main memory before
|
|
|
|
* the completion flgas are set.
|
|
|
|
*/
|
|
|
|
MB();
|
|
|
|
|
|
|
|
/* signal children that they may read the result data */
|
|
|
|
my_ctl_pointer->flag=tag;
|
|
|
|
|
|
|
|
/* copy data to user supplied buffer */
|
- 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
|
|
|
rc=ompi_datatype_copy_content_same_ddt(dtype, count_this_stripe,
|
2008-04-02 23:02:33 +04:00
|
|
|
(char *)buf+dt_extent*count_processed,
|
|
|
|
(char *)my_data_pointer);
|
|
|
|
if( 0 != rc ) {
|
|
|
|
return OMPI_ERROR;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* "free" the shared-memory working buffer */
|
|
|
|
rc=free_sm2_shared_buffer(sm_module);
|
|
|
|
if( OMPI_SUCCESS != rc ) {
|
|
|
|
goto Error;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* update the count of elements processed */
|
|
|
|
count_processed+=count_this_stripe;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* return */
|
|
|
|
return rc;
|
|
|
|
|
|
|
|
Error:
|
|
|
|
return rc;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Shared memory blocking broadcast.
|
|
|
|
*/
|
|
|
|
int mca_coll_sm2_bcast_intra(void *buf, int count,
|
|
|
|
struct ompi_datatype_t *dtype, int root,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module)
|
2008-04-02 23:02:33 +04:00
|
|
|
{
|
|
|
|
/* local variables */
|
|
|
|
int rc;
|
|
|
|
|
|
|
|
rc= mca_coll_sm2_fanout(buf, count, dtype, root, comm, module);
|
|
|
|
if( OMPI_SUCCESS != rc ) {
|
|
|
|
goto Error;
|
|
|
|
}
|
|
|
|
|
|
|
|
return OMPI_SUCCESS;
|
|
|
|
|
|
|
|
Error:
|
|
|
|
return rc;
|
|
|
|
}
|