ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
8b2d723fd4
openmpi/ompi/mca/coll/tuned/coll_tuned.h

557 строки
30 KiB
C
Исходник Обычный вид История

Add simple algorithms to support MPI_IN_PLACE for MPI_Alltoall, MPI_Alltoallv, and MPI_Alltoallw. Working on faster algorithms for tuned that will come at a later time. cmr=v1.7.3:ticket=trac:2965 This commit was SVN r28952. The following Trac tickets were found above: Ticket 2965 --> https://svn.open-mpi.org/trac/ompi/ticket/2965
2013-07-25 23:19:41 +04:00
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
snapshot as I move machines do not use This commit was SVN r7103.
2005-08-31 05:43:48 +04:00
/*
Merging in the jjhursey-ft-cr-stable branch (r13912 : HEAD). This merge adds Checkpoint/Restart support to Open MPI. The initial frameworks and components support a LAM/MPI-like implementation. This commit follows the risk assessment presented to the Open MPI core development group on Feb. 22, 2007. This commit closes trac:158 More details to follow. This commit was SVN r14051. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r13912 The following Trac tickets were found above: Ticket 158 --> https://svn.open-mpi.org/trac/ompi/ticket/158
2007-03-17 02:11:45 +03:00
* Copyright (c) 2004-2007 The Trustees of Indiana University and Indiana
Update the copyright notices for IU and UTK. This commit was SVN r7999.
2005-11-05 22:57:48 +03:00
* University Research and Technology
* Corporation. All rights reserved.
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
* Copyright (c) 2004-2009 The University of Tennessee and The University
Update the copyright notices for IU and UTK. This commit was SVN r7999.
2005-11-05 22:57:48 +03:00
* of Tennessee Research Foundation. All rights
* reserved.
snapshot as I move machines do not use This commit was SVN r7103.
2005-08-31 05:43:48 +04:00
* Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
* University of Stuttgart. All rights reserved.
* Copyright (c) 2004-2005 The Regents of the University of California.
* All rights reserved.
Add MPI_Alltoallv to tuned collectives and add a pairwise implementation of MPI_Alltoallv. However, do not change the default behavior for now. The only way to use new pairwise implementation is via mca parameters. This commit was SVN r18394.
2008-05-07 06:31:24 +04:00
* Copyright (c) 2008 Sun Microsystems, Inc. All rights reserved.
Fixes trac:1392, #1400 * add "register" function to mca_base_component_t * converted coll:basic and paffinity:linux and paffinity:solaris to use this function * we'll convert the rest over time (I'll file a ticket once all this is committed) * add 32 bytes of "reserved" space to the end of mca_base_component_t and mca_base_component_data_2_0_0_t to make future upgrades [slightly] easier * new mca_base_component_t size: 196 bytes * new mca_base_component_data_2_0_0_t size: 36 bytes * MCA base version bumped to v2.0 * '''We now refuse to load components that are not MCA v2.0.x''' * all MCA frameworks versions bumped to v2.0 * be a little more explicit about version numbers in the MCA base * add big comment in mca.h about versioning philosophy This commit was SVN r19073. The following Trac tickets were found above: Ticket 1392 --> https://svn.open-mpi.org/trac/ompi/ticket/1392
2008-07-29 02:40:57 +04:00
* Copyright (c) 2008 Cisco Systems, Inc. All rights reserved.
Add simple algorithms to support MPI_IN_PLACE for MPI_Alltoall, MPI_Alltoallv, and MPI_Alltoallw. Working on faster algorithms for tuned that will come at a later time. cmr=v1.7.3:ticket=trac:2965 This commit was SVN r28952. The following Trac tickets were found above: Ticket 2965 --> https://svn.open-mpi.org/trac/ompi/ticket/2965
2013-07-25 23:19:41 +04:00
* Copyright (c) 2013 Los Alamos National Security, LLC. All rights
* reserved.
snapshot as I move machines do not use This commit was SVN r7103.
2005-08-31 05:43:48 +04:00
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* $HEADER$
*/
#ifndef MCA_COLL_TUNED_EXPORT_H
#define MCA_COLL_TUNED_EXPORT_H
#include "ompi_config.h"
#include "mpi.h"
Next step in the project split, mainly source code re-arranging - move files out of toplevel include/ and etc/, moving it into the sub-projects - rather than including config headers with <project>/include, have them as <project> - require all headers to be included with a project prefix, with the exception of the config headers ({opal,orte,ompi}_config.h mpi.h, and mpif.h) This commit was SVN r8985.
2006-02-12 04:33:29 +03:00
#include "opal/mca/mca.h"
#include "ompi/mca/coll/coll.h"
#include "ompi/request/request.h"
snapshot as I move machines do not use This commit was SVN r7103.
2005-08-31 05:43:48 +04:00
/* need to include our own topo prototypes so we can malloc data on the comm correctly */
#include "coll_tuned_topo.h"
snapshot while switching systems but, dynamic rules from a user defined config file is almost there now This commit was SVN r7943.
2005-11-01 03:19:05 +03:00
/* also need the dynamic rule structures */
#include "coll_tuned_dynamic_rules.h"
/* some fixed value index vars to simplify certain operations */
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
typedef enum COLLTYPE {
ALLGATHER = 0, /* 0 */
ALLGATHERV, /* 1 */
ALLREDUCE, /* 2 */
ALLTOALL, /* 3 */
ALLTOALLV, /* 4 */
ALLTOALLW, /* 5 */
BARRIER, /* 6 */
BCAST, /* 7 */
EXSCAN, /* 8 */
GATHER, /* 9 */
GATHERV, /* 10 */
REDUCE, /* 11 */
REDUCESCATTER, /* 12 */
SCAN, /* 13 */
SCATTER, /* 14 */
SCATTERV, /* 15 */
COLLCOUNT /* 16 end counter keep it as last element */
} COLLTYPE_T;
snapshot while switching systems but, dynamic rules from a user defined config file is almost there now This commit was SVN r7943.
2005-11-01 03:19:05 +03:00
cleaned up ready for changes to move cached data off the MCW module to the component (where they belong) This commit was SVN r8407.
2005-12-08 06:14:57 +03:00
/* defined arg lists to simply auto inclusion of user overriding decision functions */
Fixes trac:1392, #1400 * add "register" function to mca_base_component_t * converted coll:basic and paffinity:linux and paffinity:solaris to use this function * we'll convert the rest over time (I'll file a ticket once all this is committed) * add 32 bytes of "reserved" space to the end of mca_base_component_t and mca_base_component_data_2_0_0_t to make future upgrades [slightly] easier * new mca_base_component_t size: 196 bytes * new mca_base_component_data_2_0_0_t size: 36 bytes * MCA base version bumped to v2.0 * '''We now refuse to load components that are not MCA v2.0.x''' * all MCA frameworks versions bumped to v2.0 * be a little more explicit about version numbers in the MCA base * add big comment in mca.h about versioning philosophy This commit was SVN r19073. The following Trac tickets were found above: Ticket 1392 --> https://svn.open-mpi.org/trac/ompi/ticket/1392
2008-07-29 02:40:57 +04:00
#define ALLGATHER_ARGS void *sbuf, int scount, struct ompi_datatype_t *sdtype, void *rbuf, int rcount, struct ompi_datatype_t *rdtype, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define ALLGATHERV_ARGS void *sbuf, int scount, struct ompi_datatype_t *sdtype, void * rbuf, int *rcounts, int *disps, struct ompi_datatype_t *rdtype, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define ALLREDUCE_ARGS void *sbuf, void *rbuf, int count, struct ompi_datatype_t *dtype, struct ompi_op_t *op, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define ALLTOALL_ARGS void *sbuf, int scount, struct ompi_datatype_t *sdtype, void* rbuf, int rcount, struct ompi_datatype_t *rdtype, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define ALLTOALLV_ARGS void *sbuf, int *scounts, int *sdisps, struct ompi_datatype_t *sdtype, void *rbuf, int *rcounts, int *rdisps, struct ompi_datatype_t *rdtype, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define ALLTOALLW_ARGS void *sbuf, int *scounts, int *sdisps, struct ompi_datatype_t **sdtypes, void *rbuf, int *rcounts, int *rdisps, struct ompi_datatype_t **rdtypes, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define BARRIER_ARGS struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define BCAST_ARGS void *buff, int count, struct ompi_datatype_t *datatype, int root, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define EXSCAN_ARGS void *sbuf, void *rbuf, int count, struct ompi_datatype_t *dtype, struct ompi_op_t *op, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define GATHER_ARGS void *sbuf, int scount, struct ompi_datatype_t *sdtype, void *rbuf, int rcount, struct ompi_datatype_t *rdtype, int root, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define GATHERV_ARGS void *sbuf, int scount, struct ompi_datatype_t *sdtype, void *rbuf, int *rcounts, int *disps, struct ompi_datatype_t *rdtype, int root, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define REDUCE_ARGS void *sbuf, void* rbuf, int count, struct ompi_datatype_t *dtype, struct ompi_op_t *op, int root, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define REDUCESCATTER_ARGS void *sbuf, void *rbuf, int *rcounts, struct ompi_datatype_t *dtype, struct ompi_op_t *op, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define SCAN_ARGS void *sbuf, void *rbuf, int count, struct ompi_datatype_t *dtype, struct ompi_op_t *op, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define SCATTER_ARGS void *sbuf, int scount, struct ompi_datatype_t *sdtype, void *rbuf, int rcount, struct ompi_datatype_t *rdtype, int root, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
#define SCATTERV_ARGS void *sbuf, int *scounts, int *disps, struct ompi_datatype_t *sdtype, void* rbuf, int rcount, struct ompi_datatype_t *rdtype, int root, struct ompi_communicator_t *comm, mca_coll_base_module_t *module
cleaned up ready for changes to move cached data off the MCW module to the component (where they belong) This commit was SVN r8407.
2005-12-08 06:14:57 +03:00
/* end defined arg lists to simply auto inclusion of user overriding decision functions */
snapshot while switching systems but, dynamic rules from a user defined config file is almost there now This commit was SVN r7943.
2005-11-01 03:19:05 +03:00
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
BEGIN_C_DECLS
/* these are the same across all modules and are loaded at component query time */
extern int ompi_coll_tuned_stream;
extern int ompi_coll_tuned_priority;
extern int ompi_coll_tuned_preallocate_memory_comm_size_limit;
MCA/base: Add new MCA variable system Features: - Support for an override parameter file (openmpi-mca-param-override.conf). Variable values in this file can not be overridden by any file or environment value. - Support for boolean, unsigned, and unsigned long long variables. - Support for true/false values. - Support for enumerations on integer variables. - Support for MPIT scope, verbosity, and binding. - Support for command line source. - Support for setting variable source via the environment using OMPI_MCA_SOURCE_<var name>=source (either command or file:filename) - Cleaner API. - Support for variable groups (equivalent to MPIT categories). Notes: - Variables must be created with a backing store (char **, int *, or bool *) that must live at least as long as the variable. - Creating a variable with the MCA_BASE_VAR_FLAG_SETTABLE enables the use of mca_base_var_set_value() to change the value. - String values are duplicated when the variable is registered. It is up to the caller to free the original value if necessary. The new value will be freed by the mca_base_var system and must not be freed by the user. - Variables with constant scope may not be settable. - Variable groups (and all associated variables) are deregistered when the component is closed or the component repository item is freed. This prevents a segmentation fault from accessing a variable after its component is unloaded. - After some discussion we decided we should remove the automatic registration of component priority variables. Few component actually made use of this feature. - The enumerator interface was updated to be general enough to handle future uses of the interface. - The code to generate ompi_info output has been moved into the MCA variable system. See mca_base_var_dump(). opal: update core and components to mca_base_var system orte: update core and components to mca_base_var system ompi: update core and components to mca_base_var system This commit also modifies the rmaps framework. The following variables were moved from ppr and lama: rmaps_base_pernode, rmaps_base_n_pernode, rmaps_base_n_persocket. Both lama and ppr create synonyms for these variables. This commit was SVN r28236.
2013-03-28 01:09:41 +04:00
extern bool ompi_coll_tuned_use_dynamic_rules;
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
extern char* ompi_coll_tuned_dynamic_rules_filename;
extern int ompi_coll_tuned_init_tree_fanout;
extern int ompi_coll_tuned_init_chain_fanout;
extern int ompi_coll_tuned_init_max_requests;
/* forced algorithm choices */
/* this structure is for storing the indexes to the forced algorithm mca params... */
/* we get these at component query (so that registered values appear in ompi_infoi) */
struct coll_tuned_force_algorithm_mca_param_indices_t {
int algorithm_param_index; /* which algorithm you want to force */
int segsize_param_index; /* segsize to use (if supported), 0 = no segmentation */
int tree_fanout_param_index; /* tree fanout/in to use */
int chain_fanout_param_index; /* K-chain fanout/in to use */
int max_requests_param_index; /* Maximum number of outstanding send or recv requests */
};
typedef struct coll_tuned_force_algorithm_mca_param_indices_t coll_tuned_force_algorithm_mca_param_indices_t;
/* the following type is for storing actual value obtained from the MCA on each tuned module */
/* via their mca param indices lookup in the component */
/* this structure is stored once per collective type per communicator... */
struct coll_tuned_force_algorithm_params_t {
int algorithm; /* which algorithm you want to force */
int segsize; /* segsize to use (if supported), 0 = no segmentation */
int tree_fanout; /* tree fanout/in to use */
int chain_fanout; /* K-chain fanout/in to use */
int max_requests; /* Maximum number of outstanding send or recv requests */
};
typedef struct coll_tuned_force_algorithm_params_t coll_tuned_force_algorithm_params_t;
/* the indices to the MCA params so that modules can look them up at open / comm create time */
extern coll_tuned_force_algorithm_mca_param_indices_t ompi_coll_tuned_forced_params[COLLCOUNT];
/* the actual max algorithm values (readonly), loaded at component open */
extern int ompi_coll_tuned_forced_max_algorithms[COLLCOUNT];
/*
* coll API functions
*/
/* API functions */
int ompi_coll_tuned_init_query(bool enable_progress_threads,
bool enable_mpi_threads);
mca_coll_base_module_t *
ompi_coll_tuned_comm_query(struct ompi_communicator_t *comm, int *priority);
/* API functions of decision functions and any implementations */
/*
* Note this gets long as we have to have a prototype for each
* MPI collective 4 times.. 2 for the comm type and 2 for each decision
* type.
* we might cut down the decision prototypes by conditional compiling
*/
/* All Gather */
int ompi_coll_tuned_allgather_intra_dec_fixed(ALLGATHER_ARGS);
int ompi_coll_tuned_allgather_intra_dec_dynamic(ALLGATHER_ARGS);
int ompi_coll_tuned_allgather_intra_do_forced(ALLGATHER_ARGS);
int ompi_coll_tuned_allgather_intra_do_this(ALLGATHER_ARGS, int algorithm, int faninout, int segsize);
int ompi_coll_tuned_allgather_intra_check_forced_init(coll_tuned_force_algorithm_mca_param_indices_t *mca_param_indices);
int ompi_coll_tuned_allgather_intra_bruck(ALLGATHER_ARGS);
int ompi_coll_tuned_allgather_intra_recursivedoubling(ALLGATHER_ARGS);
int ompi_coll_tuned_allgather_intra_ring(ALLGATHER_ARGS);
int ompi_coll_tuned_allgather_intra_neighborexchange(ALLGATHER_ARGS);
int ompi_coll_tuned_allgather_intra_basic_linear(ALLGATHER_ARGS);
int ompi_coll_tuned_allgather_intra_two_procs(ALLGATHER_ARGS);
int ompi_coll_tuned_allgather_inter_dec_fixed(ALLGATHER_ARGS);
int ompi_coll_tuned_allgather_inter_dec_dynamic(ALLGATHER_ARGS);
/* All GatherV */
int ompi_coll_tuned_allgatherv_intra_dec_fixed(ALLGATHERV_ARGS);
int ompi_coll_tuned_allgatherv_intra_dec_dynamic(ALLGATHERV_ARGS);
int ompi_coll_tuned_allgatherv_intra_do_forced(ALLGATHERV_ARGS);
int ompi_coll_tuned_allgatherv_intra_do_this(ALLGATHERV_ARGS, int algorithm, int faninout, int segsize);
int ompi_coll_tuned_allgatherv_intra_check_forced_init(coll_tuned_force_algorithm_mca_param_indices_t *mca_param_indices);
int ompi_coll_tuned_allgatherv_intra_bruck(ALLGATHERV_ARGS);
int ompi_coll_tuned_allgatherv_intra_ring(ALLGATHERV_ARGS);
int ompi_coll_tuned_allgatherv_intra_neighborexchange(ALLGATHERV_ARGS);
int ompi_coll_tuned_allgatherv_intra_basic_default(ALLGATHERV_ARGS);
int ompi_coll_tuned_allgatherv_intra_two_procs(ALLGATHERV_ARGS);
int ompi_coll_tuned_allgatherv_inter_dec_fixed(ALLGATHERV_ARGS);
int ompi_coll_tuned_allgatherv_inter_dec_dynamic(ALLGATHERV_ARGS);
/* All Reduce */
int ompi_coll_tuned_allreduce_intra_dec_fixed(ALLREDUCE_ARGS);
int ompi_coll_tuned_allreduce_intra_dec_dynamic(ALLREDUCE_ARGS);
int ompi_coll_tuned_allreduce_intra_do_forced(ALLREDUCE_ARGS);
int ompi_coll_tuned_allreduce_intra_do_this(ALLREDUCE_ARGS, int algorithm, int faninout, int segsize);
int ompi_coll_tuned_allreduce_intra_check_forced_init (coll_tuned_force_algorithm_mca_param_indices_t *mca_param_indices);
int ompi_coll_tuned_allreduce_intra_nonoverlapping(ALLREDUCE_ARGS);
int ompi_coll_tuned_allreduce_intra_recursivedoubling(ALLREDUCE_ARGS);
int ompi_coll_tuned_allreduce_intra_ring(ALLREDUCE_ARGS);
int ompi_coll_tuned_allreduce_intra_ring_segmented(ALLREDUCE_ARGS, uint32_t segsize);
int ompi_coll_tuned_allreduce_intra_basic_linear(ALLREDUCE_ARGS);
int ompi_coll_tuned_allreduce_inter_dec_fixed(ALLREDUCE_ARGS);
int ompi_coll_tuned_allreduce_inter_dec_dynamic(ALLREDUCE_ARGS);
/* AlltoAll */
int ompi_coll_tuned_alltoall_intra_dec_fixed(ALLTOALL_ARGS);
int ompi_coll_tuned_alltoall_intra_dec_dynamic(ALLTOALL_ARGS);
int ompi_coll_tuned_alltoall_intra_do_forced(ALLTOALL_ARGS);
int ompi_coll_tuned_alltoall_intra_do_this(ALLTOALL_ARGS, int algorithm, int faninout, int segsize, int max_requests);
int ompi_coll_tuned_alltoall_intra_check_forced_init (coll_tuned_force_algorithm_mca_param_indices_t *mca_param_indices);
int ompi_coll_tuned_alltoall_intra_pairwise(ALLTOALL_ARGS);
int ompi_coll_tuned_alltoall_intra_bruck(ALLTOALL_ARGS);
int ompi_coll_tuned_alltoall_intra_basic_linear(ALLTOALL_ARGS);
int ompi_coll_tuned_alltoall_intra_linear_sync(ALLTOALL_ARGS, int max_requests);
int ompi_coll_tuned_alltoall_intra_two_procs(ALLTOALL_ARGS);
int ompi_coll_tuned_alltoall_inter_dec_fixed(ALLTOALL_ARGS);
int ompi_coll_tuned_alltoall_inter_dec_dynamic(ALLTOALL_ARGS);
/* AlltoAllV */
int ompi_coll_tuned_alltoallv_intra_dec_fixed(ALLTOALLV_ARGS);
int ompi_coll_tuned_alltoallv_intra_dec_dynamic(ALLTOALLV_ARGS);
int ompi_coll_tuned_alltoallv_intra_do_forced(ALLTOALLV_ARGS);
int ompi_coll_tuned_alltoallv_intra_do_this(ALLTOALLV_ARGS, int algorithm);
int ompi_coll_tuned_alltoallv_intra_check_forced_init(coll_tuned_force_algorithm_mca_param_indices_t *mca_param_indices);
int ompi_coll_tuned_alltoallv_intra_pairwise(ALLTOALLV_ARGS);
int ompi_coll_tuned_alltoallv_intra_basic_linear(ALLTOALLV_ARGS);
int ompi_coll_tuned_alltoallv_inter_dec_fixed(ALLTOALLV_ARGS);
int ompi_coll_tuned_alltoallv_inter_dec_dynamic(ALLTOALLV_ARGS);
/* AlltoAllW */
int ompi_coll_tuned_alltoallw_intra_dec_fixed(ALLTOALLW_ARGS);
int ompi_coll_tuned_alltoallw_intra_dec_dynamic(ALLTOALLW_ARGS);
int ompi_coll_tuned_alltoallw_inter_dec_fixed(ALLTOALLW_ARGS);
int ompi_coll_tuned_alltoallw_inter_dec_dynamic(ALLTOALLW_ARGS);
/* Barrier */
int ompi_coll_tuned_barrier_intra_dec_fixed(BARRIER_ARGS);
int ompi_coll_tuned_barrier_intra_dec_dynamic(BARRIER_ARGS);
int ompi_coll_tuned_barrier_intra_do_forced(BARRIER_ARGS);
int ompi_coll_tuned_barrier_intra_do_this(BARRIER_ARGS, int algorithm, int faninout, int segsize);
int ompi_coll_tuned_barrier_intra_check_forced_init (coll_tuned_force_algorithm_mca_param_indices_t *mca_param_indices);
int ompi_coll_tuned_barrier_inter_dec_fixed(BARRIER_ARGS);
int ompi_coll_tuned_barrier_inter_dec_dynamic(BARRIER_ARGS);
int ompi_coll_tuned_barrier_intra_doublering(BARRIER_ARGS);
int ompi_coll_tuned_barrier_intra_recursivedoubling(BARRIER_ARGS);
int ompi_coll_tuned_barrier_intra_bruck(BARRIER_ARGS);
int ompi_coll_tuned_barrier_intra_two_procs(BARRIER_ARGS);
int ompi_coll_tuned_barrier_intra_linear(BARRIER_ARGS);
int ompi_coll_tuned_barrier_intra_tree(BARRIER_ARGS);
/* Bcast */
int ompi_coll_tuned_bcast_intra_generic( BCAST_ARGS, uint32_t count_by_segment, ompi_coll_tree_t* tree );
int ompi_coll_tuned_bcast_intra_dec_fixed(BCAST_ARGS);
int ompi_coll_tuned_bcast_intra_dec_dynamic(BCAST_ARGS);
int ompi_coll_tuned_bcast_intra_do_forced(BCAST_ARGS);
int ompi_coll_tuned_bcast_intra_do_this(BCAST_ARGS, int algorithm, int faninout, int segsize);
int ompi_coll_tuned_bcast_intra_check_forced_init (coll_tuned_force_algorithm_mca_param_indices_t *mca_param_indices);
int ompi_coll_tuned_bcast_intra_basic_linear(BCAST_ARGS);
int ompi_coll_tuned_bcast_intra_chain(BCAST_ARGS, uint32_t segsize, int32_t chains);
int ompi_coll_tuned_bcast_intra_pipeline(BCAST_ARGS, uint32_t segsize);
int ompi_coll_tuned_bcast_intra_binomial(BCAST_ARGS, uint32_t segsize);
int ompi_coll_tuned_bcast_intra_bintree(BCAST_ARGS, uint32_t segsize);
int ompi_coll_tuned_bcast_intra_split_bintree(BCAST_ARGS, uint32_t segsize);
int ompi_coll_tuned_bcast_inter_dec_fixed(BCAST_ARGS);
int ompi_coll_tuned_bcast_inter_dec_dynamic(BCAST_ARGS);
/* Exscan */
int ompi_coll_tuned_exscan_intra_dec_fixed(EXSCAN_ARGS);
int ompi_coll_tuned_exscan_intra_dec_dynamic(EXSCAN_ARGS);
int ompi_coll_tuned_exscan_inter_dec_fixed(EXSCAN_ARGS);
int ompi_coll_tuned_exscan_inter_dec_dynamic(EXSCAN_ARGS);
/* Gather */
int ompi_coll_tuned_gather_intra_dec_fixed(GATHER_ARGS);
int ompi_coll_tuned_gather_intra_dec_dynamic(GATHER_ARGS);
int ompi_coll_tuned_gather_intra_do_forced(GATHER_ARGS);
int ompi_coll_tuned_gather_intra_do_this(GATHER_ARGS, int algorithm, int faninout, int segsize);
int ompi_coll_tuned_gather_intra_check_forced_init (coll_tuned_force_algorithm_mca_param_indices_t *mca_param_indices);
int ompi_coll_tuned_gather_intra_basic_linear(GATHER_ARGS);
int ompi_coll_tuned_gather_intra_binomial(GATHER_ARGS);
int ompi_coll_tuned_gather_intra_linear_sync(GATHER_ARGS, int first_segment_size);
int ompi_coll_tuned_gather_inter_dec_fixed(GATHER_ARGS);
int ompi_coll_tuned_gather_inter_dec_dynamic(GATHER_ARGS);
/* GatherV */
int ompi_coll_tuned_gatherv_intra_dec_fixed(GATHERV_ARGS);
int ompi_coll_tuned_gatherv_intra_dec_dynamic(GATHER_ARGS);
int ompi_coll_tuned_gatherv_inter_dec_fixed(GATHER_ARGS);
int ompi_coll_tuned_gatherv_inter_dec_dynamic(GATHER_ARGS);
/* Reduce */
int ompi_coll_tuned_reduce_generic( REDUCE_ARGS, ompi_coll_tree_t* tree, int count_by_segment, int max_outstanding_reqs );
int ompi_coll_tuned_reduce_intra_dec_fixed(REDUCE_ARGS);
int ompi_coll_tuned_reduce_intra_dec_dynamic(REDUCE_ARGS);
int ompi_coll_tuned_reduce_intra_do_forced(REDUCE_ARGS);
int ompi_coll_tuned_reduce_intra_do_this(REDUCE_ARGS, int algorithm, int faninout, int segsize, int max_oustanding_reqs);
int ompi_coll_tuned_reduce_intra_check_forced_init (coll_tuned_force_algorithm_mca_param_indices_t *mca_param_indices);
int ompi_coll_tuned_reduce_intra_basic_linear(REDUCE_ARGS);
int ompi_coll_tuned_reduce_intra_chain(REDUCE_ARGS, uint32_t segsize, int fanout, int max_outstanding_reqs );
int ompi_coll_tuned_reduce_intra_pipeline(REDUCE_ARGS, uint32_t segsize, int max_outstanding_reqs );
int ompi_coll_tuned_reduce_intra_binary(REDUCE_ARGS, uint32_t segsize, int max_outstanding_reqs );
int ompi_coll_tuned_reduce_intra_binomial(REDUCE_ARGS, uint32_t segsize, int max_outstanding_reqs );
int ompi_coll_tuned_reduce_intra_in_order_binary(REDUCE_ARGS, uint32_t segsize, int max_outstanding_reqs );
int ompi_coll_tuned_reduce_inter_dec_fixed(REDUCE_ARGS);
int ompi_coll_tuned_reduce_inter_dec_dynamic(REDUCE_ARGS);
/* Reduce_scatter */
int ompi_coll_tuned_reduce_scatter_intra_dec_fixed(REDUCESCATTER_ARGS);
int ompi_coll_tuned_reduce_scatter_intra_dec_dynamic(REDUCESCATTER_ARGS);
int ompi_coll_tuned_reduce_scatter_intra_do_forced(REDUCESCATTER_ARGS);
int ompi_coll_tuned_reduce_scatter_intra_do_this(REDUCESCATTER_ARGS, int algorithm, int faninout, int segsize);
int ompi_coll_tuned_reduce_scatter_intra_check_forced_init (coll_tuned_force_algorithm_mca_param_indices_t *mca_param_indices);
int ompi_coll_tuned_reduce_scatter_intra_nonoverlapping(REDUCESCATTER_ARGS);
int ompi_coll_tuned_reduce_scatter_intra_basic_recursivehalving(REDUCESCATTER_ARGS);
int ompi_coll_tuned_reduce_scatter_intra_ring(REDUCESCATTER_ARGS);
int ompi_coll_tuned_reduce_scatter_inter_dec_fixed(REDUCESCATTER_ARGS);
int ompi_coll_tuned_reduce_scatter_inter_dec_dynamic(REDUCESCATTER_ARGS);
/* Scan */
int ompi_coll_tuned_scan_intra_dec_fixed(SCAN_ARGS);
int ompi_coll_tuned_scan_intra_dec_dynamic(SCAN_ARGS);
int ompi_coll_tuned_scan_inter_dec_fixed(SCAN_ARGS);
int ompi_coll_tuned_scan_inter_dec_dynamic(SCAN_ARGS);
/* Scatter */
int ompi_coll_tuned_scatter_intra_dec_fixed(SCATTER_ARGS);
int ompi_coll_tuned_scatter_intra_dec_dynamic(SCATTER_ARGS);
int ompi_coll_tuned_scatter_intra_do_forced(SCATTER_ARGS);
int ompi_coll_tuned_scatter_intra_do_this(SCATTER_ARGS, int algorithm, int faninout, int segsize);
int ompi_coll_tuned_scatter_intra_check_forced_init (coll_tuned_force_algorithm_mca_param_indices_t *mca_param_indices);
int ompi_coll_tuned_scatter_intra_basic_linear(SCATTER_ARGS);
int ompi_coll_tuned_scatter_intra_binomial(SCATTER_ARGS);
int ompi_coll_tuned_scatter_inter_dec_fixed(SCATTER_ARGS);
int ompi_coll_tuned_scatter_inter_dec_dynamic(SCATTER_ARGS);
/* ScatterV */
int ompi_coll_tuned_scatterv_intra_dec_fixed(SCATTERV_ARGS);
int ompi_coll_tuned_scatterv_intra_dec_dynamic(SCATTERV_ARGS);
int ompi_coll_tuned_scatterv_inter_dec_fixed(SCATTERV_ARGS);
int ompi_coll_tuned_scatterv_inter_dec_dynamic(SCATTERV_ARGS);
int mca_coll_tuned_ft_event(int state);
/* Utility functions */
static inline void ompi_coll_tuned_free_reqs(ompi_request_t **reqs, int count)
{
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
int i;
for (i = 0; i < count; ++i)
ompi_request_free(&reqs[i]);
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
}
snapshot as I move machines do not use This commit was SVN r7103.
2005-08-31 05:43:48 +04:00
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
struct mca_coll_tuned_component_t {
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
/** Base coll component */
Fixes trac:1392, #1400 * add "register" function to mca_base_component_t * converted coll:basic and paffinity:linux and paffinity:solaris to use this function * we'll convert the rest over time (I'll file a ticket once all this is committed) * add 32 bytes of "reserved" space to the end of mca_base_component_t and mca_base_component_data_2_0_0_t to make future upgrades [slightly] easier * new mca_base_component_t size: 196 bytes * new mca_base_component_data_2_0_0_t size: 36 bytes * MCA base version bumped to v2.0 * '''We now refuse to load components that are not MCA v2.0.x''' * all MCA frameworks versions bumped to v2.0 * be a little more explicit about version numbers in the MCA base * add big comment in mca.h about versioning philosophy This commit was SVN r19073. The following Trac tickets were found above: Ticket 1392 --> https://svn.open-mpi.org/trac/ompi/ticket/1392
2008-07-29 02:40:57 +04:00
mca_coll_base_component_2_0_0_t super;
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
/** MCA parameter: Priority of this component */
int tuned_priority;
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
/** global stuff that I need the component to store */
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
/* MCA parameters first */
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
/* cached decision table stuff (moved from MCW module) */
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
ompi_coll_alg_rule_t *all_base_rules;
};
/**
* Convenience typedef
*/
typedef struct mca_coll_tuned_component_t mca_coll_tuned_component_t;
/**
* Global component instance
*/
OMPI_MODULE_DECLSPEC extern mca_coll_tuned_component_t mca_coll_tuned_component;
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
/*
* Data structure for hanging data off the communicator
* i.e. per module instance
*/
struct mca_coll_tuned_comm_t {
/* standard data for requests and PML usage */
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
/* Precreate space for requests
* Note this does not effect basic,
* but if in wrong context can confuse a debugger
* this is controlled by an MCA param
*/
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
ompi_request_t **mcct_reqs;
int mcct_num_reqs;
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
/*
* tuned topo information caching per communicator
*
* for each communicator we cache the topo information so we can
* reuse without regenerating if we change the root, [or fanout]
* then regenerate and recache this information
*/
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
/* general tree with n fan out */
ompi_coll_tree_t *cached_ntree;
int cached_ntree_root;
int cached_ntree_fanout;
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
/* binary tree */
ompi_coll_tree_t *cached_bintree;
int cached_bintree_root;
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
/* binomial tree */
ompi_coll_tree_t *cached_bmtree;
int cached_bmtree_root;
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
/* binomial tree */
ompi_coll_tree_t *cached_in_order_bmtree;
int cached_in_order_bmtree_root;
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
/* chained tree (fanout followed by pipelines) */
ompi_coll_tree_t *cached_chain;
int cached_chain_root;
int cached_chain_fanout;
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
/* pipeline */
ompi_coll_tree_t *cached_pipeline;
int cached_pipeline_root;
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
/* in-order binary tree (root of the in-order binary tree is rank 0) */
ompi_coll_tree_t *cached_in_order_bintree;
/* moving to the component */
ompi_coll_com_rule_t *com_rules[COLLCOUNT]; /* the communicator rules for each MPI collective for ONLY my comsize */
/* for forced algorithms we store the information on the module */
/* previously we only had one shared copy, ops, it really is per comm/module */
coll_tuned_force_algorithm_params_t user_forced[COLLCOUNT];
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
};
typedef struct mca_coll_tuned_comm_t mca_coll_tuned_comm_t;
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
struct mca_coll_tuned_module_t {
Fixes trac:1392, #1400 * add "register" function to mca_base_component_t * converted coll:basic and paffinity:linux and paffinity:solaris to use this function * we'll convert the rest over time (I'll file a ticket once all this is committed) * add 32 bytes of "reserved" space to the end of mca_base_component_t and mca_base_component_data_2_0_0_t to make future upgrades [slightly] easier * new mca_base_component_t size: 196 bytes * new mca_base_component_data_2_0_0_t size: 36 bytes * MCA base version bumped to v2.0 * '''We now refuse to load components that are not MCA v2.0.x''' * all MCA frameworks versions bumped to v2.0 * be a little more explicit about version numbers in the MCA base * add big comment in mca.h about versioning philosophy This commit was SVN r19073. The following Trac tickets were found above: Ticket 1392 --> https://svn.open-mpi.org/trac/ompi/ticket/1392
2008-07-29 02:40:57 +04:00
mca_coll_base_module_t super;
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
mca_coll_tuned_comm_t *tuned_data;
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
};
typedef struct mca_coll_tuned_module_t mca_coll_tuned_module_t;
OBJ_CLASS_DECLARATION(mca_coll_tuned_module_t);
Start moving stuff out of modules that should be in the component. This commit was SVN r8874.
2006-02-01 23:50:14 +03:00
Add simple algorithms to support MPI_IN_PLACE for MPI_Alltoall, MPI_Alltoallv, and MPI_Alltoallw. Working on faster algorithms for tuned that will come at a later time. cmr=v1.7.3:ticket=trac:2965 This commit was SVN r28952. The following Trac tickets were found above: Ticket 2965 --> https://svn.open-mpi.org/trac/ompi/ticket/2965
2013-07-25 23:19:41 +04:00
static inline void mca_coll_tuned_free_reqs(ompi_request_t ** reqs,
int count)
{
int i;
for (i = 0; i < count; ++i)
ompi_request_free(reqs + i);
}
Rework the selection logic for the tuned collectives. All supported collectives now are able to use the dynamic rules. Moreover, these rules are loaded only once, and stored at the component level. All communicators are able to use these rules (not only MPI_COMM_WORLD as until now). A lot of minor corrections, memory management issues and reduction in the amount of memory used by the tuned collectives. This commit was SVN r21825.
2009-08-15 01:06:23 +04:00
END_C_DECLS
Big collective commit. I lightly test it, but I think it should be quite stable. Anyway, the default decision functions (for broadcast, reduce and barrier) are based on a high performance network (not TCP). It should give good performance (really good) for any network having the following caracteristics: small latency (5 microseconds) and good bandwidth (more than 1Gb/s). + Cleanup of the reduce algorithms, plus 2 new algorithms (binary and binomial). Now most of the reduce algorithms use a generic tree based function for completing the reduce. + Added macros for computing the trees (they are used for bcast and reduce right now). + Allow the usage of all 5 topologies. + Jelena's implementation of a binary tree that can be used for non commutative operations. Right now only the tree building function is there, it will get activated soon. + Some others minor cleanups. This commit was SVN r12326.
2006-10-27 02:53:05 +04:00
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
#define COLL_TUNED_UPDATE_BINTREE( OMPI_COMM, TUNED_MODULE, ROOT ) \
Big collective commit. I lightly test it, but I think it should be quite stable. Anyway, the default decision functions (for broadcast, reduce and barrier) are based on a high performance network (not TCP). It should give good performance (really good) for any network having the following caracteristics: small latency (5 microseconds) and good bandwidth (more than 1Gb/s). + Cleanup of the reduce algorithms, plus 2 new algorithms (binary and binomial). Now most of the reduce algorithms use a generic tree based function for completing the reduce. + Added macros for computing the trees (they are used for bcast and reduce right now). + Allow the usage of all 5 topologies. + Jelena's implementation of a binary tree that can be used for non commutative operations. Right now only the tree building function is there, it will get activated soon. + Some others minor cleanups. This commit was SVN r12326.
2006-10-27 02:53:05 +04:00
do { \
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
mca_coll_tuned_comm_t* coll_comm = (TUNED_MODULE)->tuned_data; \
Big collective commit. I lightly test it, but I think it should be quite stable. Anyway, the default decision functions (for broadcast, reduce and barrier) are based on a high performance network (not TCP). It should give good performance (really good) for any network having the following caracteristics: small latency (5 microseconds) and good bandwidth (more than 1Gb/s). + Cleanup of the reduce algorithms, plus 2 new algorithms (binary and binomial). Now most of the reduce algorithms use a generic tree based function for completing the reduce. + Added macros for computing the trees (they are used for bcast and reduce right now). + Allow the usage of all 5 topologies. + Jelena's implementation of a binary tree that can be used for non commutative operations. Right now only the tree building function is there, it will get activated soon. + Some others minor cleanups. This commit was SVN r12326.
2006-10-27 02:53:05 +04:00
if( !( (coll_comm->cached_bintree) \
&& (coll_comm->cached_bintree_root == (ROOT)) ) ) { \
if( coll_comm->cached_bintree ) { /* destroy previous binomial if defined */ \
ompi_coll_tuned_topo_destroy_tree( &(coll_comm->cached_bintree) ); \
} \
coll_comm->cached_bintree = ompi_coll_tuned_topo_build_tree(2,(OMPI_COMM),(ROOT)); \
coll_comm->cached_bintree_root = (ROOT); \
} \
} while (0)
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
#define COLL_TUNED_UPDATE_BMTREE( OMPI_COMM, TUNED_MODULE, ROOT ) \
Big collective commit. I lightly test it, but I think it should be quite stable. Anyway, the default decision functions (for broadcast, reduce and barrier) are based on a high performance network (not TCP). It should give good performance (really good) for any network having the following caracteristics: small latency (5 microseconds) and good bandwidth (more than 1Gb/s). + Cleanup of the reduce algorithms, plus 2 new algorithms (binary and binomial). Now most of the reduce algorithms use a generic tree based function for completing the reduce. + Added macros for computing the trees (they are used for bcast and reduce right now). + Allow the usage of all 5 topologies. + Jelena's implementation of a binary tree that can be used for non commutative operations. Right now only the tree building function is there, it will get activated soon. + Some others minor cleanups. This commit was SVN r12326.
2006-10-27 02:53:05 +04:00
do { \
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
mca_coll_tuned_comm_t* coll_comm = (TUNED_MODULE)->tuned_data; \
Big collective commit. I lightly test it, but I think it should be quite stable. Anyway, the default decision functions (for broadcast, reduce and barrier) are based on a high performance network (not TCP). It should give good performance (really good) for any network having the following caracteristics: small latency (5 microseconds) and good bandwidth (more than 1Gb/s). + Cleanup of the reduce algorithms, plus 2 new algorithms (binary and binomial). Now most of the reduce algorithms use a generic tree based function for completing the reduce. + Added macros for computing the trees (they are used for bcast and reduce right now). + Allow the usage of all 5 topologies. + Jelena's implementation of a binary tree that can be used for non commutative operations. Right now only the tree building function is there, it will get activated soon. + Some others minor cleanups. This commit was SVN r12326.
2006-10-27 02:53:05 +04:00
if( !( (coll_comm->cached_bmtree) \
&& (coll_comm->cached_bmtree_root == (ROOT)) ) ) { \
if( coll_comm->cached_bmtree ) { /* destroy previous binomial if defined */ \
ompi_coll_tuned_topo_destroy_tree( &(coll_comm->cached_bmtree) ); \
} \
coll_comm->cached_bmtree = ompi_coll_tuned_topo_build_bmtree( (OMPI_COMM), (ROOT) ); \
coll_comm->cached_bmtree_root = (ROOT); \
} \
} while (0)
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
#define COLL_TUNED_UPDATE_IN_ORDER_BMTREE( OMPI_COMM, TUNED_MODULE, ROOT ) \
added binomial tree based Gather alogrithm, passed IBM and Intel tests This commit was SVN r13835.
2007-02-28 04:11:01 +03:00
do { \
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
mca_coll_tuned_comm_t* coll_comm = (TUNED_MODULE)->tuned_data; \
added binomial tree based Gather alogrithm, passed IBM and Intel tests This commit was SVN r13835.
2007-02-28 04:11:01 +03:00
if( !( (coll_comm->cached_in_order_bmtree) \
&& (coll_comm->cached_in_order_bmtree_root == (ROOT)) ) ) { \
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
if( coll_comm->cached_in_order_bmtree ) { /* destroy previous binomial if defined */ \
ompi_coll_tuned_topo_destroy_tree( &(coll_comm->cached_in_order_bmtree) ); \
added binomial tree based Gather alogrithm, passed IBM and Intel tests This commit was SVN r13835.
2007-02-28 04:11:01 +03:00
} \
coll_comm->cached_in_order_bmtree = ompi_coll_tuned_topo_build_in_order_bmtree( (OMPI_COMM), (ROOT) ); \
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
coll_comm->cached_in_order_bmtree_root = (ROOT); \
added binomial tree based Gather alogrithm, passed IBM and Intel tests This commit was SVN r13835.
2007-02-28 04:11:01 +03:00
} \
} while (0)
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
#define COLL_TUNED_UPDATE_PIPELINE( OMPI_COMM, TUNED_MODULE, ROOT ) \
Big collective commit. I lightly test it, but I think it should be quite stable. Anyway, the default decision functions (for broadcast, reduce and barrier) are based on a high performance network (not TCP). It should give good performance (really good) for any network having the following caracteristics: small latency (5 microseconds) and good bandwidth (more than 1Gb/s). + Cleanup of the reduce algorithms, plus 2 new algorithms (binary and binomial). Now most of the reduce algorithms use a generic tree based function for completing the reduce. + Added macros for computing the trees (they are used for bcast and reduce right now). + Allow the usage of all 5 topologies. + Jelena's implementation of a binary tree that can be used for non commutative operations. Right now only the tree building function is there, it will get activated soon. + Some others minor cleanups. This commit was SVN r12326.
2006-10-27 02:53:05 +04:00
do { \
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
mca_coll_tuned_comm_t* coll_comm = (TUNED_MODULE)->tuned_data; \
Big collective commit. I lightly test it, but I think it should be quite stable. Anyway, the default decision functions (for broadcast, reduce and barrier) are based on a high performance network (not TCP). It should give good performance (really good) for any network having the following caracteristics: small latency (5 microseconds) and good bandwidth (more than 1Gb/s). + Cleanup of the reduce algorithms, plus 2 new algorithms (binary and binomial). Now most of the reduce algorithms use a generic tree based function for completing the reduce. + Added macros for computing the trees (they are used for bcast and reduce right now). + Allow the usage of all 5 topologies. + Jelena's implementation of a binary tree that can be used for non commutative operations. Right now only the tree building function is there, it will get activated soon. + Some others minor cleanups. This commit was SVN r12326.
2006-10-27 02:53:05 +04:00
if( !( (coll_comm->cached_pipeline) \
&& (coll_comm->cached_pipeline_root == (ROOT)) ) ) { \
if (coll_comm->cached_pipeline) { /* destroy previous pipeline if defined */ \
ompi_coll_tuned_topo_destroy_tree( &(coll_comm->cached_pipeline) ); \
} \
coll_comm->cached_pipeline = ompi_coll_tuned_topo_build_chain( 1, (OMPI_COMM), (ROOT) ); \
coll_comm->cached_pipeline_root = (ROOT); \
} \
} while (0)
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
#define COLL_TUNED_UPDATE_CHAIN( OMPI_COMM, TUNED_MODULE, ROOT, FANOUT ) \
Big collective commit. I lightly test it, but I think it should be quite stable. Anyway, the default decision functions (for broadcast, reduce and barrier) are based on a high performance network (not TCP). It should give good performance (really good) for any network having the following caracteristics: small latency (5 microseconds) and good bandwidth (more than 1Gb/s). + Cleanup of the reduce algorithms, plus 2 new algorithms (binary and binomial). Now most of the reduce algorithms use a generic tree based function for completing the reduce. + Added macros for computing the trees (they are used for bcast and reduce right now). + Allow the usage of all 5 topologies. + Jelena's implementation of a binary tree that can be used for non commutative operations. Right now only the tree building function is there, it will get activated soon. + Some others minor cleanups. This commit was SVN r12326.
2006-10-27 02:53:05 +04:00
do { \
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
mca_coll_tuned_comm_t* coll_comm = (TUNED_MODULE)->tuned_data; \
Big collective commit. I lightly test it, but I think it should be quite stable. Anyway, the default decision functions (for broadcast, reduce and barrier) are based on a high performance network (not TCP). It should give good performance (really good) for any network having the following caracteristics: small latency (5 microseconds) and good bandwidth (more than 1Gb/s). + Cleanup of the reduce algorithms, plus 2 new algorithms (binary and binomial). Now most of the reduce algorithms use a generic tree based function for completing the reduce. + Added macros for computing the trees (they are used for bcast and reduce right now). + Allow the usage of all 5 topologies. + Jelena's implementation of a binary tree that can be used for non commutative operations. Right now only the tree building function is there, it will get activated soon. + Some others minor cleanups. This commit was SVN r12326.
2006-10-27 02:53:05 +04:00
if( !( (coll_comm->cached_chain) \
&& (coll_comm->cached_chain_root == (ROOT)) \
&& (coll_comm->cached_chain_fanout == (FANOUT)) ) ) { \
if( coll_comm->cached_chain) { /* destroy previous chain if defined */ \
ompi_coll_tuned_topo_destroy_tree( &(coll_comm->cached_chain) ); \
} \
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
coll_comm->cached_chain = ompi_coll_tuned_topo_build_chain((FANOUT), (OMPI_COMM), (ROOT)); \
Big collective commit. I lightly test it, but I think it should be quite stable. Anyway, the default decision functions (for broadcast, reduce and barrier) are based on a high performance network (not TCP). It should give good performance (really good) for any network having the following caracteristics: small latency (5 microseconds) and good bandwidth (more than 1Gb/s). + Cleanup of the reduce algorithms, plus 2 new algorithms (binary and binomial). Now most of the reduce algorithms use a generic tree based function for completing the reduce. + Added macros for computing the trees (they are used for bcast and reduce right now). + Allow the usage of all 5 topologies. + Jelena's implementation of a binary tree that can be used for non commutative operations. Right now only the tree building function is there, it will get activated soon. + Some others minor cleanups. This commit was SVN r12326.
2006-10-27 02:53:05 +04:00
coll_comm->cached_chain_root = (ROOT); \
coll_comm->cached_chain_fanout = (FANOUT); \
} \
Adding segmented binary reduce algorithm which works with non-commutative operations. Implementation passed intel: MPI_Reduce_c , MPI_Reduce_loc_c, and MPI_Reduce_user_c tests over TCP, BTL MX, and MTL MX, as well as, mpi_test_suite Reduce tests (up to 64 nodes). The algorithm is still not activated by decision function (will be in the near future). This commit was SVN r13657.
2007-02-15 01:38:38 +03:00
} while (0)
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
#define COLL_TUNED_UPDATE_IN_ORDER_BINTREE( OMPI_COMM, TUNED_MODULE ) \
Adding segmented binary reduce algorithm which works with non-commutative operations. Implementation passed intel: MPI_Reduce_c , MPI_Reduce_loc_c, and MPI_Reduce_user_c tests over TCP, BTL MX, and MTL MX, as well as, mpi_test_suite Reduce tests (up to 64 nodes). The algorithm is still not activated by decision function (will be in the near future). This commit was SVN r13657.
2007-02-15 01:38:38 +03:00
do { \
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
mca_coll_tuned_comm_t* coll_comm = (TUNED_MODULE)->tuned_data; \
Adding segmented binary reduce algorithm which works with non-commutative operations. Implementation passed intel: MPI_Reduce_c , MPI_Reduce_loc_c, and MPI_Reduce_user_c tests over TCP, BTL MX, and MTL MX, as well as, mpi_test_suite Reduce tests (up to 64 nodes). The algorithm is still not activated by decision function (will be in the near future). This commit was SVN r13657.
2007-02-15 01:38:38 +03:00
if( !(coll_comm->cached_in_order_bintree) ) { \
/* In-order binary tree topology is defined by communicator size */ \
/* Thus, there is no need to destroy anything */ \
coll_comm->cached_in_order_bintree = \
Update collectives selection logic to allow for multiple components to be used at nce (up to one unique collective module per collective function). Matches r15795:15921 of the tmp/bwb-coll-select branch This commit was SVN r15924. The following SVN revisions from the original message are invalid or inconsistent and therefore were not cross-referenced: r15795 r15921
2007-08-19 07:37:49 +04:00
ompi_coll_tuned_topo_build_in_order_bintree((OMPI_COMM)); \
Adding segmented binary reduce algorithm which works with non-commutative operations. Implementation passed intel: MPI_Reduce_c , MPI_Reduce_loc_c, and MPI_Reduce_user_c tests over TCP, BTL MX, and MTL MX, as well as, mpi_test_suite Reduce tests (up to 64 nodes). The algorithm is still not activated by decision function (will be in the near future). This commit was SVN r13657.
2007-02-15 01:38:38 +03:00
} \
Big collective commit. I lightly test it, but I think it should be quite stable. Anyway, the default decision functions (for broadcast, reduce and barrier) are based on a high performance network (not TCP). It should give good performance (really good) for any network having the following caracteristics: small latency (5 microseconds) and good bandwidth (more than 1Gb/s). + Cleanup of the reduce algorithms, plus 2 new algorithms (binary and binomial). Now most of the reduce algorithms use a generic tree based function for completing the reduce. + Added macros for computing the trees (they are used for bcast and reduce right now). + Allow the usage of all 5 topologies. + Jelena's implementation of a binary tree that can be used for non commutative operations. Right now only the tree building function is there, it will get activated soon. + Some others minor cleanups. This commit was SVN r12326.
2006-10-27 02:53:05 +04:00
} while (0)
Updates & upgrades: - consistent arguments checking (not allowing to select an algorithm which is not available) - consistent way of computing the segcount (number of datatypes by segment). - small cleanups. - more informative debugging messages. This commit was SVN r12545.
2006-11-10 22:54:09 +03:00
/**
* This macro give a generic way to compute the best count of
* the segment (i.e. the number of complete datatypes that
* can fit in the specified SEGSIZE). Beware, when this macro
* is called, the SEGCOUNT should be initialized to the count as
* expected by the collective call.
*/
#define COLL_TUNED_COMPUTED_SEGCOUNT(SEGSIZE, TYPELNG, SEGCOUNT) \
Correctly compute the number of elements in a segment. For broadcast send the correct size for all intermediary nodes. This commit was SVN r12552.
2006-11-11 02:04:50 +03:00
if( ((SEGSIZE) >= (TYPELNG)) && \
((SEGSIZE) < ((TYPELNG) * (SEGCOUNT))) ) { \
Updates & upgrades: - consistent arguments checking (not allowing to select an algorithm which is not available) - consistent way of computing the segcount (number of datatypes by segment). - small cleanups. - more informative debugging messages. This commit was SVN r12545.
2006-11-10 22:54:09 +03:00
size_t residual; \
(SEGCOUNT) = (int)((SEGSIZE) / (TYPELNG)); \
residual = (SEGSIZE) - (SEGCOUNT) * (TYPELNG); \
if( residual > ((TYPELNG) >> 1) ) \
(SEGCOUNT)++; \
} \
Modification to allreduce ring algorithm: - the block sizes are computed in more uniformn way. The first k blocks may be 1 element larger than the remaining blocks. The algorithm passed Intel Allreduce_c and Allreduce_loc_c tests, and IMB-3.2 Allreduce, over TCP and both btl and mtl MX (up to 128 processes). The algorithm still only supports commutative operations. This commit was SVN r13738.
2007-02-21 22:30:08 +03:00
/**
* This macro gives a generic wait to compute the well distributed block counts
* when the count and number of blocks are fixed.
* Macro returns "early-block" count, "late-block" count, and "split-index"
* which is the block at which we switch from "early-block" count to
* the "late-block" count.
* count = split_index * early_block_count +
* (block_count - split_index) * late_block_count
* We do not perform ANY error checks - make sure that the input values
* make sense (eg. count > num_blocks).
*/
#define COLL_TUNED_COMPUTE_BLOCKCOUNT( COUNT, NUM_BLOCKS, SPLIT_INDEX, \
EARLY_BLOCK_COUNT, LATE_BLOCK_COUNT ) \
EARLY_BLOCK_COUNT = LATE_BLOCK_COUNT = COUNT / NUM_BLOCKS; \
SPLIT_INDEX = COUNT % NUM_BLOCKS; \
if (0 != SPLIT_INDEX) { \
EARLY_BLOCK_COUNT = EARLY_BLOCK_COUNT + 1; \
} \
snapshot as I move machines do not use This commit was SVN r7103.
2005-08-31 05:43:48 +04:00
#endif /* MCA_COLL_TUNED_EXPORT_H */
Modification to allreduce ring algorithm: - the block sizes are computed in more uniformn way. The first k blocks may be 1 element larger than the remaining blocks. The algorithm passed Intel Allreduce_c and Allreduce_loc_c tests, and IMB-3.2 Allreduce, over TCP and both btl and mtl MX (up to 128 processes). The algorithm still only supports commutative operations. This commit was SVN r13738.
2007-02-21 22:30:08 +03:00
Копировать постоянную ссылку