src/smpi/colls/smpi_automatic_selector.cpp

   1 /* Copyright (c) 2013-2017. The SimGrid Team.
   2  * All rights reserved.                                                     */
   3
   4 /* This program is free software; you can redistribute it and/or modify it
   5  * under the terms of the license (GNU LGPL) which comes with this package. */
   6
   7 #include <cfloat>
   8
   9 #include <exception>
  10
  11 #include "colls_private.h"
  12 #include "smpi_process.hpp"
  13
  14
  15 //attempt to do a quick autotuning version of the collective,
  16
  17 #define TRACE_AUTO_COLL(cat)                                                                                           \
  18   if (TRACE_is_enabled()) {                                                                                            \
  19     type_t type = s_type::s_type_get_or_null(#cat, PJ_type_get_root());                                                       \
  20     if (not type) {                                                                                                    \
  21       type = PJ_type_event_new(#cat, PJ_type_get_root());                                                              \
  22     }                                                                                                                  \
  23     char cont_name[25];                                                                                                \
  24     snprintf(cont_name, 25, "rank-%d", smpi_process()->index());                                                       \
  25     value* val = value::get_or_new(Colls::mpi_coll_##cat##_description[i].name, "1.0 1.0 1.0", type);                  \
  26     new NewEvent(SIMIX_get_clock(), PJ_container_get(cont_name), type, val);                                           \
  27   }
  28
  29 #define AUTOMATIC_COLL_BENCH(cat, ret, args, args2)                                                                    \
  30   ret Coll_##cat##_automatic::cat(COLL_UNPAREN args)                                                                   \
  31   {                                                                                                                    \
  32     double time1, time2, time_min = DBL_MAX;                                                                           \
  33     int min_coll = -1, global_coll = -1;                                                                               \
  34     int i;                                                                                                             \
  35     double buf_in, buf_out, max_min = DBL_MAX;                                                                         \
  36     for (i = 0; Colls::mpi_coll_##cat##_description[i].name; i++) {                                                    \
  37       if (not strcmp(Colls::mpi_coll_##cat##_description[i].name, "automatic"))                                        \
  38         continue;                                                                                                      \
  39       if (not strcmp(Colls::mpi_coll_##cat##_description[i].name, "default"))                                          \
  40         continue;                                                                                                      \
  41       Coll_barrier_default::barrier(comm);                                                                             \
  42       TRACE_AUTO_COLL(cat)                                                                                             \
  43       time1 = SIMIX_get_clock();                                                                                       \
  44       try {                                                                                                            \
  45         ((int(*) args)Colls::mpi_coll_##cat##_description[i].coll) args2;                                              \
  46       } catch (std::exception & ex) {                                                                                  \
  47         continue;                                                                                                      \
  48       }                                                                                                                \
  49       time2   = SIMIX_get_clock();                                                                                     \
  50       buf_out = time2 - time1;                                                                                         \
  51       Coll_reduce_default::reduce((void*)&buf_out, (void*)&buf_in, 1, MPI_DOUBLE, MPI_MAX, 0, comm);                   \
  52       if (time2 - time1 < time_min) {                                                                                  \
  53         min_coll = i;                                                                                                  \
  54         time_min = time2 - time1;                                                                                      \
  55       }                                                                                                                \
  56       if (comm->rank() == 0) {                                                                                         \
  57         if (buf_in < max_min) {                                                                                        \
  58           max_min     = buf_in;                                                                                        \
  59           global_coll = i;                                                                                             \
  60         }                                                                                                              \
  61       }                                                                                                                \
  62     }                                                                                                                  \
  63     if (comm->rank() == 0) {                                                                                           \
  64       XBT_WARN("For rank 0, the quickest was %s : %f , but global was %s : %f at max",                                 \
  65                Colls::mpi_coll_##cat##_description[min_coll].name, time_min,                                           \
  66                Colls::mpi_coll_##cat##_description[global_coll].name, max_min);                                        \
  67     } else                                                                                                             \
  68       XBT_WARN("The quickest %s was %s on rank %d and took %f", #cat,                                                  \
  69                Colls::mpi_coll_##cat##_description[min_coll].name, comm->rank(), time_min);                            \
  70     return (min_coll != -1) ? MPI_SUCCESS : MPI_ERR_INTERN;                                                            \
  71   }
  72
  73 namespace simgrid{
  74 namespace smpi{
  75
  76 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLGATHERV_SIG, (send_buff, send_count, send_type, recv_buff, recv_count, recv_disps, recv_type, comm));
  77 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLREDUCE_SIG, (sbuf, rbuf, rcount, dtype, op, comm));
  78 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_GATHER_SIG, (send_buff, send_count, send_type, recv_buff, recv_count, recv_type, root, comm));
  79 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLGATHER_SIG, (send_buff,send_count,send_type,recv_buff,recv_count,recv_type,comm));
  80 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLTOALL_SIG,(send_buff, send_count, send_type, recv_buff, recv_count, recv_type,comm));
  81 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLTOALLV_SIG, (send_buff, send_counts, send_disps, send_type, recv_buff, recv_counts, recv_disps, recv_type, comm));
  82 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_BCAST_SIG , (buf, count, datatype, root, comm));
  83 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_REDUCE_SIG,(buf,rbuf, count, datatype, op, root, comm));
  84 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_REDUCE_SCATTER_SIG ,(sbuf,rbuf, rcounts,dtype,op,comm));
  85 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_SCATTER_SIG ,(sendbuf, sendcount, sendtype,recvbuf, recvcount, recvtype,root, comm));
  86 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_BARRIER_SIG,(comm));
  87
  88 }
  89 }