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 #include <exception>
   9
  10 #include "colls_private.hpp"
  11 #include "smpi_process.hpp"
  12
  13 //attempt to do a quick autotuning version of the collective,
  14 #define TRACE_AUTO_COLL(cat)                                                                                           \
  15   if (TRACE_is_enabled()) {                                                                                            \
  16     simgrid::instr::EventType* type = simgrid::instr::Container::getRoot()->type_->getOrCreateEventType(#cat);         \
  17                                                                                                                        \
  18     std::string cont_name = std::string("rank-" + std::to_string(smpi_process()->index()));                            \
  19     type->addEntityValue(Colls::mpi_coll_##cat##_description[i].name, "1.0 1.0 1.0");                                  \
  20     new simgrid::instr::NewEvent(SIMIX_get_clock(), simgrid::instr::Container::byName(cont_name), type,                \
  21                                  type->getEntityValue(Colls::mpi_coll_##cat##_description[i].name));                   \
  22   }
  23
  24 #define AUTOMATIC_COLL_BENCH(cat, ret, args, args2)                                                                    \
  25   ret Coll_##cat##_automatic::cat(COLL_UNPAREN args)                                                                   \
  26   {                                                                                                                    \
  27     double time1, time2, time_min = DBL_MAX;                                                                           \
  28     int min_coll = -1, global_coll = -1;                                                                               \
  29     int i;                                                                                                             \
  30     double buf_in, buf_out, max_min = DBL_MAX;                                                                         \
  31     for (i = 0; Colls::mpi_coll_##cat##_description[i].name; i++) {                                                    \
  32       if (not strcmp(Colls::mpi_coll_##cat##_description[i].name, "automatic"))                                        \
  33         continue;                                                                                                      \
  34       if (not strcmp(Colls::mpi_coll_##cat##_description[i].name, "default"))                                          \
  35         continue;                                                                                                      \
  36       Coll_barrier_default::barrier(comm);                                                                             \
  37       TRACE_AUTO_COLL(cat)                                                                                             \
  38       time1 = SIMIX_get_clock();                                                                                       \
  39       try {                                                                                                            \
  40         ((int(*) args)Colls::mpi_coll_##cat##_description[i].coll) args2;                                              \
  41       } catch (std::exception & ex) {                                                                                  \
  42         continue;                                                                                                      \
  43       }                                                                                                                \
  44       time2   = SIMIX_get_clock();                                                                                     \
  45       buf_out = time2 - time1;                                                                                         \
  46       Coll_reduce_default::reduce((void*)&buf_out, (void*)&buf_in, 1, MPI_DOUBLE, MPI_MAX, 0, comm);                   \
  47       if (time2 - time1 < time_min) {                                                                                  \
  48         min_coll = i;                                                                                                  \
  49         time_min = time2 - time1;                                                                                      \
  50       }                                                                                                                \
  51       if (comm->rank() == 0) {                                                                                         \
  52         if (buf_in < max_min) {                                                                                        \
  53           max_min     = buf_in;                                                                                        \
  54           global_coll = i;                                                                                             \
  55         }                                                                                                              \
  56       }                                                                                                                \
  57     }                                                                                                                  \
  58     if (comm->rank() == 0) {                                                                                           \
  59       XBT_WARN("For rank 0, the quickest was %s : %f , but global was %s : %f at max",                                 \
  60                Colls::mpi_coll_##cat##_description[min_coll].name, time_min,                                           \
  61                Colls::mpi_coll_##cat##_description[global_coll].name, max_min);                                        \
  62     } else                                                                                                             \
  63       XBT_WARN("The quickest %s was %s on rank %d and took %f", #cat,                                                  \
  64                Colls::mpi_coll_##cat##_description[min_coll].name, comm->rank(), time_min);                            \
  65     return (min_coll != -1) ? MPI_SUCCESS : MPI_ERR_INTERN;                                                            \
  66   }
  67
  68 namespace simgrid{
  69 namespace smpi{
  70
  71 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLGATHERV_SIG, (send_buff, send_count, send_type, recv_buff, recv_count, recv_disps, recv_type, comm));
  72 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLREDUCE_SIG, (sbuf, rbuf, rcount, dtype, op, comm));
  73 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_GATHER_SIG, (send_buff, send_count, send_type, recv_buff, recv_count, recv_type, root, comm));
  74 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLGATHER_SIG, (send_buff,send_count,send_type,recv_buff,recv_count,recv_type,comm));
  75 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLTOALL_SIG,(send_buff, send_count, send_type, recv_buff, recv_count, recv_type,comm));
  76 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));
  77 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_BCAST_SIG , (buf, count, datatype, root, comm));
  78 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_REDUCE_SIG,(buf,rbuf, count, datatype, op, root, comm));
  79 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_REDUCE_SCATTER_SIG ,(sbuf,rbuf, rcounts,dtype,op,comm));
  80 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_SCATTER_SIG ,(sendbuf, sendcount, sendtype,recvbuf, recvcount, recvtype,root, comm));
  81 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_BARRIER_SIG,(comm));
  82
  83 }
  84 }