2004-08-21 04:49:07 +04:00
|
|
|
/*
|
2007-03-17 02:11:45 +03:00
|
|
|
* Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
|
2005-11-05 22:57:48 +03:00
|
|
|
* University Research and Technology
|
|
|
|
* Corporation. All rights reserved.
|
2014-10-16 05:47:32 +04:00
|
|
|
* Copyright (c) 2004-2014 The University of Tennessee and The University
|
2005-11-05 22:57:48 +03:00
|
|
|
* of Tennessee Research Foundation. All rights
|
|
|
|
* reserved.
|
2004-11-28 23:09:25 +03:00
|
|
|
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
|
|
|
|
* University of Stuttgart. All rights reserved.
|
2005-03-24 15:43:37 +03:00
|
|
|
* Copyright (c) 2004-2005 The Regents of the University of California.
|
|
|
|
* All rights reserved.
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
* Copyright (c) 2008-2009 Cisco Systems, Inc. All rights reserved.
|
2004-11-22 04:38:40 +03:00
|
|
|
* $COPYRIGHT$
|
|
|
|
*
|
|
|
|
* Additional copyrights may follow
|
|
|
|
*
|
2004-11-22 03:37:56 +03:00
|
|
|
* $HEADER$
|
2004-08-21 04:49:07 +04:00
|
|
|
*/
|
2005-09-07 01:41:55 +04:00
|
|
|
/** @file */
|
2004-08-21 04:49:07 +04:00
|
|
|
|
|
|
|
#ifndef MCA_COLL_SM_EXPORT_H
|
|
|
|
#define MCA_COLL_SM_EXPORT_H
|
|
|
|
|
|
|
|
#include "ompi_config.h"
|
|
|
|
|
|
|
|
#include "mpi.h"
|
2015-03-06 06:50:44 +03:00
|
|
|
#include "ompi/mca/mca.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 "opal/datatype/opal_convertor.h"
|
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
|
|
|
#include "opal/mca/common/sm/common_sm.h"
|
2005-07-16 00:01:35 +04:00
|
|
|
#include "ompi/mca/coll/coll.h"
|
2004-08-21 04:49:07 +04:00
|
|
|
|
2007-08-19 07:37:49 +04:00
|
|
|
BEGIN_C_DECLS
|
|
|
|
|
2009-09-22 02:20:44 +04:00
|
|
|
/* Attempt to give some sort of progress / fairness if we're blocked
|
|
|
|
in an sm collective for a long time: call opal_progress once in a
|
|
|
|
great while. Use a "goto" label for expdiency to exit loops. */
|
|
|
|
#define SPIN_CONDITION_MAX 100000
|
|
|
|
#define SPIN_CONDITION(cond, exit_label) \
|
|
|
|
do { int i; \
|
|
|
|
if (cond) goto exit_label; \
|
|
|
|
for (i = 0; i < SPIN_CONDITION_MAX; ++i) { \
|
|
|
|
if (cond) { goto exit_label; } \
|
|
|
|
} \
|
|
|
|
opal_progress(); \
|
|
|
|
} while (1); \
|
|
|
|
exit_label:
|
2004-08-21 04:49:07 +04:00
|
|
|
|
2005-07-16 00:01:35 +04:00
|
|
|
/**
|
|
|
|
* Structure to hold the sm coll component. First it holds the
|
|
|
|
* base coll component, and then holds a bunch of
|
|
|
|
* sm-coll-component-specific stuff (e.g., current MCA param
|
|
|
|
* values).
|
|
|
|
*/
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
typedef struct mca_coll_sm_component_t {
|
2005-07-16 00:01:35 +04:00
|
|
|
/** Base coll component */
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_component_2_0_0_t super;
|
2005-07-16 00:01:35 +04:00
|
|
|
|
2005-08-24 01:22:00 +04:00
|
|
|
/** MCA parameter: Priority of this component */
|
2005-07-16 00:01:35 +04:00
|
|
|
int sm_priority;
|
2005-08-24 01:22:00 +04:00
|
|
|
|
|
|
|
/** MCA parameter: Length of a cache line or page (in bytes) */
|
|
|
|
int sm_control_size;
|
|
|
|
|
2005-09-03 15:49:46 +04:00
|
|
|
/** MCA parameter: Number of "in use" flags in each
|
|
|
|
communicator's area in the data mpool */
|
|
|
|
int sm_comm_num_in_use_flags;
|
|
|
|
|
2005-08-24 01:22:00 +04:00
|
|
|
/** MCA parameter: Number of segments for each communicator in
|
|
|
|
the data mpool */
|
2005-09-03 15:49:46 +04:00
|
|
|
int sm_comm_num_segments;
|
2005-08-24 01:22:00 +04:00
|
|
|
|
|
|
|
/** MCA parameter: Fragment size for data */
|
|
|
|
int sm_fragment_size;
|
2004-08-21 04:49:07 +04:00
|
|
|
|
2005-08-24 01:22:00 +04:00
|
|
|
/** MCA parameter: Degree of tree for tree-based collectives */
|
|
|
|
int sm_tree_degree;
|
2004-08-21 04:49:07 +04:00
|
|
|
|
2005-09-03 15:49:46 +04:00
|
|
|
/** MCA parameter: Number of processes to use in the
|
|
|
|
calculation of the "info" MCA parameter */
|
|
|
|
int sm_info_comm_size;
|
|
|
|
|
2005-09-07 01:41:55 +04:00
|
|
|
/******* end of MCA params ********/
|
|
|
|
|
2005-10-01 03:12:23 +04:00
|
|
|
/** How many fragment segments are protected by a single
|
|
|
|
in-use flags. This is solely so that we can only perform
|
|
|
|
the division once and then just use the value without
|
|
|
|
having to re-calculate. */
|
|
|
|
int sm_segs_per_inuse_flag;
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
} mca_coll_sm_component_t;
|
2005-08-24 01:22:00 +04:00
|
|
|
|
2005-09-02 16:57:47 +04:00
|
|
|
/**
|
|
|
|
* Structure for representing a node in the tree
|
|
|
|
*/
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
typedef struct mca_coll_sm_tree_node_t {
|
2005-09-02 16:57:47 +04:00
|
|
|
/** Arbitrary ID number, starting from 0 */
|
|
|
|
int mcstn_id;
|
|
|
|
/** Pointer to parent, or NULL if root */
|
|
|
|
struct mca_coll_sm_tree_node_t *mcstn_parent;
|
|
|
|
/** Number of children, or 0 if a leaf */
|
|
|
|
int mcstn_num_children;
|
|
|
|
/** Pointer to an array of children, or NULL if 0 ==
|
|
|
|
mcstn_num_children */
|
|
|
|
struct mca_coll_sm_tree_node_t **mcstn_children;
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
} mca_coll_sm_tree_node_t;
|
2005-09-02 16:57:47 +04:00
|
|
|
|
2005-09-07 01:41:55 +04:00
|
|
|
/**
|
|
|
|
* Simple structure comprising the "in use" flags. Contains two
|
|
|
|
* members: the number of processes that are currently using this
|
|
|
|
* set of segments and the operation number of the current
|
|
|
|
* operation.
|
|
|
|
*/
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
typedef struct mca_coll_sm_in_use_flag_t {
|
2005-09-07 01:41:55 +04:00
|
|
|
/** Number of processes currently using this set of
|
|
|
|
segments */
|
|
|
|
volatile uint32_t mcsiuf_num_procs_using;
|
|
|
|
/** Must match data->mcb_count */
|
|
|
|
volatile uint32_t mcsiuf_operation_count;
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
} mca_coll_sm_in_use_flag_t;
|
2005-09-07 01:41:55 +04:00
|
|
|
|
2005-08-24 01:22:00 +04:00
|
|
|
/**
|
|
|
|
* Structure containing pointers to various arrays of data in the
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
* per-communicator shmem data segment (one of these indexes a
|
|
|
|
* single segment in the per-communicator shmem data segment).
|
|
|
|
* Nothing is hard-coded because all the array lengths and
|
|
|
|
* displacements of the pointers all depend on how many processes
|
|
|
|
* are in the communicator.
|
2005-08-24 01:22:00 +04:00
|
|
|
*/
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
typedef struct mca_coll_sm_data_index_t {
|
2005-09-02 16:57:47 +04:00
|
|
|
/** Pointer to beginning of control data */
|
2005-10-04 18:52:59 +04:00
|
|
|
uint32_t volatile *mcbmi_control;
|
2005-09-02 16:57:47 +04:00
|
|
|
/** Pointer to beginning of message fragment data */
|
|
|
|
char *mcbmi_data;
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
} mca_coll_sm_data_index_t;
|
2005-08-24 01:22:00 +04:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Structure for the sm coll module to hang off the communicator.
|
|
|
|
* Contains communicator-specific information, including pointers
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
* into the per-communicator shmem data data segment for this
|
|
|
|
* comm's sm collective operations area.
|
2005-08-24 01:22:00 +04:00
|
|
|
*/
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
typedef struct mca_coll_sm_comm_t {
|
|
|
|
/* Meta data that we get back from the common mmap allocation
|
|
|
|
function */
|
2010-06-09 20:58:52 +04:00
|
|
|
mca_common_sm_module_t *sm_bootstrap_meta;
|
2005-08-24 01:22:00 +04:00
|
|
|
|
2005-09-02 16:57:47 +04:00
|
|
|
/** Pointer to my barrier control pages (odd index pages are
|
|
|
|
"in", even index pages are "out") */
|
|
|
|
uint32_t *mcb_barrier_control_me;
|
2005-08-24 01:22:00 +04:00
|
|
|
|
2005-09-02 16:57:47 +04:00
|
|
|
/** Pointer to my parent's barrier control pages (will be NULL
|
|
|
|
for communicator rank 0; odd index pages are "in", even
|
|
|
|
index pages are "out") */
|
|
|
|
uint32_t *mcb_barrier_control_parent;
|
|
|
|
|
|
|
|
/** Pointers to my childrens' barrier control pages (they're
|
|
|
|
contiguous in memory, so we only point to the base -- the
|
|
|
|
number of children is in my entry in the mcb_tree); will
|
|
|
|
be NULL if this process has no children (odd index pages
|
|
|
|
are "in", even index pages are "out") */
|
|
|
|
uint32_t *mcb_barrier_control_children;
|
2005-08-24 02:02:28 +04:00
|
|
|
|
2005-09-02 16:57:47 +04:00
|
|
|
/** Number of barriers that we have executed (i.e., which set
|
|
|
|
of barrier buffers to use). */
|
|
|
|
int mcb_barrier_count;
|
2005-08-24 02:02:28 +04:00
|
|
|
|
2005-09-07 01:41:55 +04:00
|
|
|
/** "In use" flags indicating which segments are available */
|
|
|
|
mca_coll_sm_in_use_flag_t *mcb_in_use_flags;
|
|
|
|
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
/** Array of indexes into the per-communicator shmem data
|
|
|
|
segment for control and data fragment passing (containing
|
|
|
|
pointers to each segments control and data areas). */
|
|
|
|
mca_coll_sm_data_index_t *mcb_data_index;
|
2005-08-24 02:02:28 +04:00
|
|
|
|
2005-09-02 16:57:47 +04:00
|
|
|
/** Array of graph nodes representing the tree used for
|
|
|
|
communications */
|
|
|
|
mca_coll_sm_tree_node_t *mcb_tree;
|
2005-08-24 02:02:28 +04:00
|
|
|
|
2005-09-02 16:57:47 +04:00
|
|
|
/** Operation number (i.e., which segment number to use) */
|
2005-09-07 01:41:55 +04:00
|
|
|
uint32_t mcb_operation_count;
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
} mca_coll_sm_comm_t;
|
2007-08-19 07:37:49 +04:00
|
|
|
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
/** Coll sm module */
|
|
|
|
typedef struct mca_coll_sm_module_t {
|
|
|
|
/** Base module */
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t super;
|
2007-08-19 07:37:49 +04:00
|
|
|
|
2009-10-02 21:13:56 +04:00
|
|
|
/* Whether this module has been lazily initialized or not yet */
|
|
|
|
bool enabled;
|
|
|
|
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
/* Data that hangs off the communicator */
|
|
|
|
mca_coll_sm_comm_t *sm_comm_data;
|
|
|
|
|
|
|
|
/* Underlying reduce function and module */
|
2007-08-19 07:37:49 +04:00
|
|
|
mca_coll_base_module_reduce_fn_t previous_reduce;
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *previous_reduce_module;
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
} mca_coll_sm_module_t;
|
2007-08-19 07:37:49 +04:00
|
|
|
OBJ_CLASS_DECLARATION(mca_coll_sm_module_t);
|
2005-01-30 04:42:57 +03:00
|
|
|
|
2005-07-16 00:01:35 +04:00
|
|
|
/**
|
|
|
|
* Global component instance
|
2005-01-30 04:42:57 +03:00
|
|
|
*/
|
2006-08-27 08:49:02 +04:00
|
|
|
OMPI_MODULE_DECLSPEC extern mca_coll_sm_component_t mca_coll_sm_component;
|
2004-08-21 04:49:07 +04:00
|
|
|
|
|
|
|
/*
|
2005-01-30 04:42:57 +03:00
|
|
|
* coll module functions
|
2004-08-21 04:49:07 +04:00
|
|
|
*/
|
2005-03-27 17:05:23 +04:00
|
|
|
int mca_coll_sm_init_query(bool enable_progress_threads,
|
2007-08-19 07:37:49 +04:00
|
|
|
bool enable_mpi_threads);
|
2004-08-21 04:49:07 +04:00
|
|
|
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *
|
2007-08-19 07:37:49 +04:00
|
|
|
mca_coll_sm_comm_query(struct ompi_communicator_t *comm, int *priority);
|
2009-10-02 21:13:56 +04:00
|
|
|
|
|
|
|
/* Lazily enable a module (since it involves expensive/slow mmap
|
|
|
|
allocation, etc.) */
|
|
|
|
int ompi_coll_sm_lazy_enable(mca_coll_base_module_t *module,
|
|
|
|
struct ompi_communicator_t *comm);
|
2004-08-21 04:49:07 +04:00
|
|
|
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_allgather_intra(void *sbuf, int scount,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *sdtype,
|
|
|
|
void *rbuf, int rcount,
|
|
|
|
struct ompi_datatype_t *rdtype,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2007-08-19 07:37:49 +04:00
|
|
|
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_allgatherv_intra(void *sbuf, int scount,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *sdtype,
|
|
|
|
void * rbuf, int *rcounts, int *disps,
|
|
|
|
struct ompi_datatype_t *rdtype,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_allreduce_intra(void *sbuf, void *rbuf, int count,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *dtype,
|
|
|
|
struct ompi_op_t *op,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_alltoall_intra(void *sbuf, int scount,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *sdtype,
|
|
|
|
void* rbuf, int rcount,
|
|
|
|
struct ompi_datatype_t *rdtype,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_alltoallv_intra(void *sbuf, int *scounts, int *sdisps,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *sdtype,
|
|
|
|
void *rbuf, int *rcounts, int *rdisps,
|
|
|
|
struct ompi_datatype_t *rdtype,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_alltoallw_intra(void *sbuf, int *scounts, int *sdisps,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t **sdtypes,
|
|
|
|
void *rbuf, int *rcounts, int *rdisps,
|
|
|
|
struct ompi_datatype_t **rdtypes,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2007-08-19 07:37:49 +04:00
|
|
|
int mca_coll_sm_barrier_intra(struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_bcast_intra(void *buff, int count,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *datatype,
|
|
|
|
int root,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_bcast_log_intra(void *buff, int count,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *datatype,
|
|
|
|
int root,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_exscan_intra(void *sbuf, void *rbuf, int count,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *dtype,
|
|
|
|
struct ompi_op_t *op,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_gather_intra(void *sbuf, int scount,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *sdtype, void *rbuf,
|
|
|
|
int rcount, struct ompi_datatype_t *rdtype,
|
|
|
|
int root, struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_gatherv_intra(void *sbuf, int scount,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *sdtype, void *rbuf,
|
|
|
|
int *rcounts, int *disps,
|
|
|
|
struct ompi_datatype_t *rdtype, int root,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_reduce_intra(void *sbuf, void* rbuf, int count,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *dtype,
|
|
|
|
struct ompi_op_t *op,
|
|
|
|
int root,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_reduce_log_intra(void *sbuf, void* rbuf, int count,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *dtype,
|
|
|
|
struct ompi_op_t *op,
|
|
|
|
int root,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_reduce_scatter_intra(void *sbuf, void *rbuf,
|
2007-08-19 07:37:49 +04:00
|
|
|
int *rcounts,
|
|
|
|
struct ompi_datatype_t *dtype,
|
|
|
|
struct ompi_op_t *op,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_scan_intra(void *sbuf, void *rbuf, int count,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *dtype,
|
|
|
|
struct ompi_op_t *op,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_scatter_intra(void *sbuf, int scount,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *sdtype, void *rbuf,
|
|
|
|
int rcount, struct ompi_datatype_t *rdtype,
|
|
|
|
int root, struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2005-01-30 04:42:57 +03:00
|
|
|
int mca_coll_sm_scatterv_intra(void *sbuf, int *scounts, int *disps,
|
2007-08-19 07:37:49 +04:00
|
|
|
struct ompi_datatype_t *sdtype,
|
|
|
|
void* rbuf, int rcount,
|
|
|
|
struct ompi_datatype_t *rdtype, int root,
|
|
|
|
struct ompi_communicator_t *comm,
|
2008-07-29 02:40:57 +04:00
|
|
|
mca_coll_base_module_t *module);
|
2007-03-17 02:11:45 +03:00
|
|
|
|
|
|
|
int mca_coll_sm_ft_event(int state);
|
2005-01-30 04:42:57 +03:00
|
|
|
|
2005-09-07 17:33:43 +04:00
|
|
|
/**
|
|
|
|
* Global variables used in the macros (essentially constants, so
|
|
|
|
* these are thread safe)
|
|
|
|
*/
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
extern uint32_t mca_coll_sm_one;
|
2005-09-07 17:33:43 +04:00
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Macro to setup flag usage
|
|
|
|
*/
|
|
|
|
#define FLAG_SETUP(flag_num, flag, data) \
|
|
|
|
(flag) = (mca_coll_sm_in_use_flag_t*) \
|
|
|
|
(((char *) (data)->mcb_in_use_flags) + \
|
2005-09-15 06:18:16 +04:00
|
|
|
((flag_num) * mca_coll_sm_component.sm_control_size))
|
2005-09-07 17:33:43 +04:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Macro to wait for the in-use flag to become idle (used by the root)
|
|
|
|
*/
|
2009-09-22 02:20:44 +04:00
|
|
|
#define FLAG_WAIT_FOR_IDLE(flag, label) \
|
|
|
|
SPIN_CONDITION(0 == (flag)->mcsiuf_num_procs_using, label)
|
2005-09-07 17:33:43 +04:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Macro to wait for a flag to indicate that it's ready for this
|
|
|
|
* operation (used by non-root processes to know when FLAG_SET() has
|
|
|
|
* been called)
|
|
|
|
*/
|
2009-09-22 02:20:44 +04:00
|
|
|
#define FLAG_WAIT_FOR_OP(flag, op, label) \
|
|
|
|
SPIN_CONDITION((op) == flag->mcsiuf_operation_count, label)
|
2005-09-07 17:33:43 +04:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Macro to set an in-use flag with relevant data to claim it
|
|
|
|
*/
|
|
|
|
#define FLAG_RETAIN(flag, num_procs, op_count) \
|
|
|
|
(flag)->mcsiuf_num_procs_using = (num_procs); \
|
2005-09-15 06:18:16 +04:00
|
|
|
(flag)->mcsiuf_operation_count = (op_count)
|
2005-09-07 17:33:43 +04:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Macro to release an in-use flag from this process
|
|
|
|
*/
|
|
|
|
#define FLAG_RELEASE(flag) \
|
2014-10-16 05:47:32 +04:00
|
|
|
(void)opal_atomic_add(&(flag)->mcsiuf_num_procs_using, -1)
|
2005-09-07 17:33:43 +04:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Macro to copy a single segment in from a user buffer to a shared
|
|
|
|
* segment
|
|
|
|
*/
|
2005-09-15 06:18:16 +04:00
|
|
|
#define COPY_FRAGMENT_IN(convertor, index, rank, iov, max_data) \
|
2005-09-07 17:33:43 +04:00
|
|
|
(iov).iov_base = \
|
|
|
|
(index)->mcbmi_data + \
|
2005-09-15 06:18:16 +04:00
|
|
|
((rank) * mca_coll_sm_component.sm_fragment_size); \
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
(iov).iov_len = (max_data); \
|
|
|
|
opal_convertor_pack(&(convertor), &(iov), &mca_coll_sm_one, \
|
2006-10-27 03:11:26 +04:00
|
|
|
&(max_data) )
|
2005-09-07 17:33:43 +04:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Macro to copy a single segment out from a shared segment to a user
|
|
|
|
* buffer
|
|
|
|
*/
|
|
|
|
#define COPY_FRAGMENT_OUT(convertor, src_rank, index, iov, max_data) \
|
|
|
|
(iov).iov_base = (((char*) (index)->mcbmi_data) + \
|
Fixes trac:1988. The little bug that turned out to be huge. Yoinks.
* Various cosmetic/style updates in the btl sm
* Clean up concept of mpool module (I think that code was written way
back when the concept of "modules" was fuzzy)
* Bring over some old fixes from the /tmp/timattox-sm-coll/ tree to
fix potential segv's when mmap'ed regions were at different
addresses in different processes (thanks Tim!).
* Change sm coll to no longer use mpool as its main source of shmem;
rather, just mmap its own segment (because it's fixed size --
there was nothing to be gained by using mpool; shedding the use of
mpool saved a lot of complexity in the sm coll setup). This
effectively made Tim's fixes moot (because now everything is an
offset into the mmap that is computed locally; there are no global
pointers). :-)
* Slightly updated common/sm to allow making mmap's for a specific
set of procs (vs. ''all'' procs in the process). This potentially
allows for same-host-inter-proc mmaps -- yay!
* Fixed many, many things in the coll sm (particularly in reduce):
* Fixed handling of MPI_IN_PLACE in reduce and allreduce
* Fixed handling of non-contiguous datatypes in reduce
* Changed the order of reductions to go from process (n-1)'s data
to process 0's data, because that's how all other OMPI coll
components work
* Fixed lots of usage of ddt functions
* When using a non-contiguous datatype, if the root process is not
(n-1), now we used a 2nd convertor to copy from shmem to the rbuf
(saves a memory copy vs. what was done before)
* Lots and lots of little cleanups, clarifications, and minor
optimizations (although still more could be done -- e.g., I think
the use of write memory barriers is fairly sub-optimal; they
could be ganged together at the root, for example)
I'm marking this as "fixes trac:1988" and closing the ticket; if something
is still broken, we can re-open the ticket.
This commit was SVN r21967.
The following Trac tickets were found above:
Ticket 1988 --> https://svn.open-mpi.org/trac/ompi/ticket/1988
2009-09-15 04:25:21 +04:00
|
|
|
((src_rank) * (mca_coll_sm_component.sm_fragment_size))); \
|
|
|
|
(iov).iov_len = (max_data); \
|
|
|
|
opal_convertor_unpack(&(convertor), &(iov), &mca_coll_sm_one, \
|
2006-10-27 03:11:26 +04:00
|
|
|
&(max_data) )
|
2005-09-07 17:33:43 +04:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Macro to memcpy a fragment between one shared segment and another
|
|
|
|
*/
|
|
|
|
#define COPY_FRAGMENT_BETWEEN(src_rank, dest_rank, index, len) \
|
|
|
|
memcpy(((index)->mcbmi_data + \
|
|
|
|
((dest_rank) * mca_coll_sm_component.sm_fragment_size)), \
|
|
|
|
((index)->mcbmi_data + \
|
|
|
|
((src_rank) * \
|
|
|
|
mca_coll_sm_component.sm_fragment_size)), \
|
2005-09-15 06:18:16 +04:00
|
|
|
(len))
|
2005-09-07 17:33:43 +04:00
|
|
|
|
|
|
|
/**
|
2005-09-15 06:18:16 +04:00
|
|
|
* Macro to tell children that a segment is ready (normalize
|
|
|
|
* the child's ID based on the shift used to calculate the "me" node
|
|
|
|
* in the tree). Used in fan out opertations.
|
2005-09-07 17:33:43 +04:00
|
|
|
*/
|
|
|
|
#define PARENT_NOTIFY_CHILDREN(children, num_children, index, value) \
|
2005-09-15 06:18:16 +04:00
|
|
|
do { \
|
|
|
|
for (i = 0; i < (num_children); ++i) { \
|
|
|
|
*((size_t*) \
|
|
|
|
(((char*) index->mcbmi_control) + \
|
|
|
|
(mca_coll_sm_component.sm_control_size * \
|
|
|
|
(((children)[i]->mcstn_id + root) % size)))) = (value); \
|
|
|
|
} \
|
|
|
|
} while (0)
|
2005-09-07 17:33:43 +04:00
|
|
|
|
|
|
|
/**
|
|
|
|
* Macro for childen to wait for parent notification (use real rank).
|
2005-09-15 06:18:16 +04:00
|
|
|
* Save the value passed and then reset it when done. Used in fan out
|
|
|
|
* operations.
|
2005-09-07 17:33:43 +04:00
|
|
|
*/
|
2009-09-22 02:20:44 +04:00
|
|
|
#define CHILD_WAIT_FOR_NOTIFY(rank, index, value, label) \
|
2005-09-15 06:18:16 +04:00
|
|
|
do { \
|
2005-10-04 18:52:59 +04:00
|
|
|
uint32_t volatile *ptr = ((uint32_t*) \
|
2005-09-15 06:18:16 +04:00
|
|
|
(((char*) index->mcbmi_control) + \
|
|
|
|
((rank) * mca_coll_sm_component.sm_control_size))); \
|
2009-09-22 02:20:44 +04:00
|
|
|
SPIN_CONDITION(0 != *ptr, label); \
|
2005-09-15 06:18:16 +04:00
|
|
|
(value) = *ptr; \
|
|
|
|
*ptr = 0; \
|
|
|
|
} while (0)
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Macro for children to tell parent that the data is ready in their
|
|
|
|
* segment. Used for fan in operations.
|
|
|
|
*/
|
|
|
|
#define CHILD_NOTIFY_PARENT(child_rank, parent_rank, index, value) \
|
2005-10-04 18:52:59 +04:00
|
|
|
((size_t volatile *) \
|
2005-09-15 06:18:16 +04:00
|
|
|
(((char*) (index)->mcbmi_control) + \
|
|
|
|
(mca_coll_sm_component.sm_control_size * \
|
|
|
|
(parent_rank))))[(child_rank)] = (value)
|
|
|
|
|
|
|
|
/**
|
|
|
|
* Macro for parent to wait for a specific child to tell it that the
|
|
|
|
* data is in the child's segment. Save the value when done. Used
|
|
|
|
* for fan in operations.
|
|
|
|
*/
|
2009-09-22 02:20:44 +04:00
|
|
|
#define PARENT_WAIT_FOR_NOTIFY_SPECIFIC(child_rank, parent_rank, index, value, label) \
|
2005-09-15 06:18:16 +04:00
|
|
|
do { \
|
2005-10-04 18:52:59 +04:00
|
|
|
size_t volatile *ptr = ((size_t volatile *) \
|
2005-09-15 06:18:16 +04:00
|
|
|
(((char*) index->mcbmi_control) + \
|
|
|
|
(mca_coll_sm_component.sm_control_size * \
|
|
|
|
(parent_rank)))) + child_rank; \
|
2009-09-22 02:20:44 +04:00
|
|
|
SPIN_CONDITION(0 != *ptr, label); \
|
2005-09-15 06:18:16 +04:00
|
|
|
(value) = *ptr; \
|
|
|
|
*ptr = 0; \
|
|
|
|
} while (0)
|
2005-09-07 17:33:43 +04:00
|
|
|
|
2007-08-19 07:37:49 +04:00
|
|
|
END_C_DECLS
|
|
|
|
|
2004-08-21 04:49:07 +04:00
|
|
|
#endif /* MCA_COLL_SM_EXPORT_H */
|