root/ompi/mca/coll/tuned/coll_tuned_reduce_scatter_decision.c

DEFINITIONS

1. ompi_coll_tuned_reduce_scatter_intra_check_forced_init
2. ompi_coll_tuned_reduce_scatter_intra_do_this

   1 /* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
   2 /*
   3  * Copyright (c) 2004-2017 The University of Tennessee and The University
   4  *                         of Tennessee Research Foundation.  All rights
   5  *                         reserved.
   6  * Copyright (c) 2015      Research Organization for Information Science
   7  *                         and Technology (RIST). All rights reserved.
   8  * $COPYRIGHT$
   9  *
  10  * Additional copyrights may follow
  11  *
  12  * $HEADER$
  13  */
  14 
  15 #include "ompi_config.h"
  16 
  17 #include "mpi.h"
  18 #include "opal/util/bit_ops.h"
  19 #include "ompi/constants.h"
  20 #include "ompi/datatype/ompi_datatype.h"
  21 #include "ompi/communicator/communicator.h"
  22 #include "ompi/mca/coll/coll.h"
  23 #include "ompi/mca/coll/base/coll_base_topo.h"
  24 #include "ompi/mca/coll/base/coll_tags.h"
  25 #include "ompi/mca/pml/pml.h"
  26 #include "ompi/op/op.h"
  27 #include "coll_tuned.h"
  28 
  29 /* reduce_scatter algorithm variables */
  30 static int coll_tuned_reduce_scatter_forced_algorithm = 0;
  31 static int coll_tuned_reduce_scatter_segment_size = 0;
  32 static int coll_tuned_reduce_scatter_tree_fanout;
  33 static int coll_tuned_reduce_scatter_chain_fanout;
  34 
  35 /* valid values for coll_tuned_reduce_scatter_forced_algorithm */
  36 static mca_base_var_enum_value_t reduce_scatter_algorithms[] = {
  37     {0, "ignore"},
  38     {1, "non-overlapping"},
  39     {2, "recursive_halving"},
  40     {3, "ring"},
  41     {4, "butterfly"},
  42     {0, NULL}
  43 };
  44 
  45 /**
  46  * The following are used by dynamic and forced rules
  47  *
  48  * publish details of each algorithm and if its forced/fixed/locked in
  49  * as you add methods/algorithms you must update this and the query/map routines
  50  *
  51  * this routine is called by the component only
  52  * this makes sure that the mca parameters are set to their initial values and
  53  * perms module does not call this they call the forced_getvalues routine
  54  * instead
  55  */
  56 
  57 int ompi_coll_tuned_reduce_scatter_intra_check_forced_init (coll_tuned_force_algorithm_mca_param_indices_t *mca_param_indices)
  58 {
  59     mca_base_var_enum_t *new_enum;
  60     int cnt;
  61 
  62     for( cnt = 0; NULL != reduce_scatter_algorithms[cnt].string; cnt++ );
  63     ompi_coll_tuned_forced_max_algorithms[REDUCESCATTER] = cnt;
  64 
  65     (void) mca_base_component_var_register(&mca_coll_tuned_component.super.collm_version,
  66                                            "reduce_scatter_algorithm_count",
  67                                            "Number of reduce_scatter algorithms available",
  68                                            MCA_BASE_VAR_TYPE_INT, NULL, 0,
  69                                            MCA_BASE_VAR_FLAG_DEFAULT_ONLY,
  70                                            OPAL_INFO_LVL_5,
  71                                            MCA_BASE_VAR_SCOPE_CONSTANT,
  72                                            &ompi_coll_tuned_forced_max_algorithms[REDUCESCATTER]);
  73 
  74     /* MPI_T: This variable should eventually be bound to a communicator */
  75     coll_tuned_reduce_scatter_forced_algorithm = 0;
  76     (void) mca_base_var_enum_create("coll_tuned_reduce_scatter_algorithms", reduce_scatter_algorithms, &new_enum);
  77     mca_param_indices->algorithm_param_index =
  78         mca_base_component_var_register(&mca_coll_tuned_component.super.collm_version,
  79                                         "reduce_scatter_algorithm",
  80                                         "Which reduce reduce_scatter algorithm is used. Can be locked down to choice of: 0 ignore, 1 non-overlapping (Reduce + Scatterv), 2 recursive halving, 3 ring, 4 butterfly",
  81                                         MCA_BASE_VAR_TYPE_INT, new_enum, 0, MCA_BASE_VAR_FLAG_SETTABLE,
  82                                         OPAL_INFO_LVL_5,
  83                                         MCA_BASE_VAR_SCOPE_ALL,
  84                                         &coll_tuned_reduce_scatter_forced_algorithm);
  85     OBJ_RELEASE(new_enum);
  86     if (mca_param_indices->algorithm_param_index < 0) {
  87         return mca_param_indices->algorithm_param_index;
  88     }
  89 
  90     coll_tuned_reduce_scatter_segment_size = 0;
  91     mca_param_indices->segsize_param_index =
  92         mca_base_component_var_register(&mca_coll_tuned_component.super.collm_version,
  93                                         "reduce_scatter_algorithm_segmentsize",
  94                                         "Segment size in bytes used by default for reduce_scatter algorithms. Only has meaning if algorithm is forced and supports segmenting. 0 bytes means no segmentation.",
  95                                         MCA_BASE_VAR_TYPE_INT, NULL, 0, MCA_BASE_VAR_FLAG_SETTABLE,
  96                                         OPAL_INFO_LVL_5,
  97                                         MCA_BASE_VAR_SCOPE_ALL,
  98                                         &coll_tuned_reduce_scatter_segment_size);
  99 
 100     coll_tuned_reduce_scatter_tree_fanout = ompi_coll_tuned_init_tree_fanout; /* get system wide default */
 101     mca_param_indices->tree_fanout_param_index =
 102         mca_base_component_var_register(&mca_coll_tuned_component.super.collm_version,
 103                                         "reduce_scatter_algorithm_tree_fanout",
 104                                         "Fanout for n-tree used for reduce_scatter algorithms. Only has meaning if algorithm is forced and supports n-tree topo based operation.",
 105                                         MCA_BASE_VAR_TYPE_INT, NULL, 0, MCA_BASE_VAR_FLAG_SETTABLE,
 106                                         OPAL_INFO_LVL_5,
 107                                         MCA_BASE_VAR_SCOPE_ALL,
 108                                         &coll_tuned_reduce_scatter_tree_fanout);
 109 
 110     coll_tuned_reduce_scatter_chain_fanout = ompi_coll_tuned_init_chain_fanout; /* get system wide default */
 111     mca_param_indices->chain_fanout_param_index =
 112       mca_base_component_var_register(&mca_coll_tuned_component.super.collm_version,
 113                                       "reduce_scatter_algorithm_chain_fanout",
 114                                       "Fanout for chains used for reduce_scatter algorithms. Only has meaning if algorithm is forced and supports chain topo based operation.",
 115                                       MCA_BASE_VAR_TYPE_INT, NULL, 0, MCA_BASE_VAR_FLAG_SETTABLE,
 116                                       OPAL_INFO_LVL_5,
 117                                       MCA_BASE_VAR_SCOPE_ALL,
 118                                       &coll_tuned_reduce_scatter_chain_fanout);
 119 
 120     return (MPI_SUCCESS);
 121 }
 122 
 123 int ompi_coll_tuned_reduce_scatter_intra_do_this(const void *sbuf, void* rbuf,
 124                                                  const int *rcounts,
 125                                                  struct ompi_datatype_t *dtype,
 126                                                  struct ompi_op_t *op,
 127                                                  struct ompi_communicator_t *comm,
 128                                                  mca_coll_base_module_t *module,
 129                                                  int algorithm, int faninout, int segsize)
 130 {
 131     OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:reduce_scatter_intra_do_this selected algorithm %d topo faninout %d segsize %d",
 132                  algorithm, faninout, segsize));
 133 
 134     switch (algorithm) {
 135     case (0): return ompi_coll_tuned_reduce_scatter_intra_dec_fixed(sbuf, rbuf, rcounts,
 136                                                                     dtype, op, comm, module);
 137     case (1): return ompi_coll_base_reduce_scatter_intra_nonoverlapping(sbuf, rbuf, rcounts,
 138                                                                         dtype, op, comm, module);
 139     case (2): return ompi_coll_base_reduce_scatter_intra_basic_recursivehalving(sbuf, rbuf, rcounts,
 140                                                                                 dtype, op, comm, module);
 141     case (3): return ompi_coll_base_reduce_scatter_intra_ring(sbuf, rbuf, rcounts,
 142                                                               dtype, op, comm, module);
 143     case (4): return ompi_coll_base_reduce_scatter_intra_butterfly(sbuf, rbuf, rcounts,
 144                                                                    dtype, op, comm, module);
 145     } /* switch */
 146     OPAL_OUTPUT((ompi_coll_tuned_stream,"coll:tuned:reduce_scatter_intra_do_this attempt to select algorithm %d when only 0-%d is valid?",
 147                  algorithm, ompi_coll_tuned_forced_max_algorithms[REDUCESCATTER]));
 148     return (MPI_ERR_ARG);
 149 }