src/smpi/colls/smpi_automatic_selector.cpp

   1 /* Copyright (c) 2013-2019. The SimGrid Team. All rights reserved.          */
   2
   3 /* This program is free software; you can redistribute it and/or modify it
   4  * under the terms of the license (GNU LGPL) which comes with this package. */
   5
   6 #include <cfloat>
   7 #include <exception>
   8
   9 #include "colls_private.hpp"
  10 #include "src/smpi/include/smpi_actor.hpp"
  11
  12 //attempt to do a quick autotuning version of the collective,
  13 #define TRACE_AUTO_COLL(cat)                                                                                           \
  14   if (TRACE_is_enabled()) {                                                                                            \
  15     simgrid::instr::EventType* type =                                                                                  \
  16         simgrid::instr::Container::get_root()->type_->by_name_or_create<simgrid::instr::EventType>(                    \
  17             _XBT_STRINGIFY(cat));                                                                                      \
  18                                                                                                                        \
  19     std::string cont_name = std::string("rank-" + std::to_string(simgrid::s4u::this_actor::get_pid()));                \
  20     type->add_entity_value(Colls::_XBT_CONCAT3(mpi_coll_, cat, _description)[i].name, "1.0 1.0 1.0");                  \
  21     new simgrid::instr::NewEvent(SIMIX_get_clock(), simgrid::instr::Container::by_name(cont_name), type,               \
  22                                  type->get_entity_value(Colls::_XBT_CONCAT3(mpi_coll_, cat, _description)[i].name));   \
  23   }
  24
  25 #define AUTOMATIC_COLL_BENCH(cat, ret, args, args2)                                                                    \
  26   ret _XBT_CONCAT2(cat, __automatic)(COLL_UNPAREN args)                                                                \
  27   {                                                                                                                    \
  28     double time1, time2, time_min = DBL_MAX;                                                                           \
  29     int min_coll = -1, global_coll = -1;                                                                               \
  30     int i;                                                                                                             \
  31     double buf_in, buf_out, max_min = DBL_MAX;                                                                         \
  32     for (i = 0; not Colls::_XBT_CONCAT3(mpi_coll_, cat, _description)[i].name.empty(); i++) {                          \
  33       if (Colls::_XBT_CONCAT3(mpi_coll_, cat, _description)[i].name == "automatic")                                    \
  34         continue;                                                                                                      \
  35       if (Colls::_XBT_CONCAT3(mpi_coll_, cat, _description)[i].name == "default")                                      \
  36         continue;                                                                                                      \
  37       barrier__default(comm);                                                                                          \
  38       TRACE_AUTO_COLL(cat)                                                                                             \
  39       time1 = SIMIX_get_clock();                                                                                       \
  40       try {                                                                                                            \
  41         ((int(*) args)Colls::_XBT_CONCAT3(mpi_coll_, cat, _description)[i].coll) args2;                                \
  42       } catch (std::exception & ex) {                                                                                  \
  43         continue;                                                                                                      \
  44       }                                                                                                                \
  45       time2   = SIMIX_get_clock();                                                                                     \
  46       buf_out = time2 - time1;                                                                                         \
  47       reduce__default((void*)&buf_out, (void*)&buf_in, 1, MPI_DOUBLE, MPI_MAX, 0, comm);                               \
  48       if (time2 - time1 < time_min) {                                                                                  \
  49         min_coll = i;                                                                                                  \
  50         time_min = time2 - time1;                                                                                      \
  51       }                                                                                                                \
  52       if (comm->rank() == 0) {                                                                                         \
  53         if (buf_in < max_min) {                                                                                        \
  54           max_min     = buf_in;                                                                                        \
  55           global_coll = i;                                                                                             \
  56         }                                                                                                              \
  57       }                                                                                                                \
  58     }                                                                                                                  \
  59     if (comm->rank() == 0) {                                                                                           \
  60       XBT_WARN("For rank 0, the quickest was %s : %f , but global was %s : %f at max",                                 \
  61                Colls::_XBT_CONCAT3(mpi_coll_, cat, _description)[min_coll].name.c_str(), time_min,                     \
  62                Colls::_XBT_CONCAT3(mpi_coll_, cat, _description)[global_coll].name.c_str(), max_min);                  \
  63     } else                                                                                                             \
  64       XBT_WARN("The quickest " _XBT_STRINGIFY(cat) " was %s on rank %d and took %f",                                   \
  65                Colls::_XBT_CONCAT3(mpi_coll_, cat, _description)[min_coll].name.c_str(), comm->rank(), time_min);      \
  66     return (min_coll != -1) ? MPI_SUCCESS : MPI_ERR_INTERN;                                                            \
  67   }
  68
  69 namespace simgrid{
  70 namespace smpi{
  71
  72 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLGATHERV_SIG, (send_buff, send_count, send_type, recv_buff, recv_count, recv_disps, recv_type, comm));
  73 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLREDUCE_SIG, (sbuf, rbuf, rcount, dtype, op, comm));
  74 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_GATHER_SIG, (send_buff, send_count, send_type, recv_buff, recv_count, recv_type, root, comm));
  75 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLGATHER_SIG, (send_buff,send_count,send_type,recv_buff,recv_count,recv_type,comm));
  76 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLTOALL_SIG,(send_buff, send_count, send_type, recv_buff, recv_count, recv_type,comm));
  77 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));
  78 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_BCAST_SIG , (buf, count, datatype, root, comm));
  79 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_REDUCE_SIG,(buf,rbuf, count, datatype, op, root, comm));
  80 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_REDUCE_SCATTER_SIG ,(sbuf,rbuf, rcounts,dtype,op,comm));
  81 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_SCATTER_SIG ,(sendbuf, sendcount, sendtype,recvbuf, recvcount, recvtype,root, comm));
  82 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_BARRIER_SIG,(comm));
  83
  84 }
  85 }