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 AUTOMATIC_COLL_BENCH(cat, ret, args, args2)                                                                    \
  14   ret _XBT_CONCAT2(cat, __automatic)(COLL_UNPAREN args)                                                                \
  15   {                                                                                                                    \
  16     double time1, time2, time_min = DBL_MAX;                                                                           \
  17     int min_coll = -1, global_coll = -1;                                                                               \
  18     int i = 0;                                                                                                         \
  19     double buf_in, buf_out, max_min = DBL_MAX;                                                                         \
  20     auto desc = simgrid::smpi::colls::get_smpi_coll_description(_XBT_STRINGIFY(cat), i);                               \
  21     while (not desc->name.empty()) {                                                                                   \
  22       if (desc->name == "automatic")                                                                                   \
  23         goto next_iteration;                                                                                           \
  24       if (desc->name == "default")                                                                                     \
  25         goto next_iteration;                                                                                           \
  26       barrier__default(comm);                                                                                          \
  27       if (TRACE_is_enabled()) {                                                                                        \
  28         simgrid::instr::EventType* type =                                                                              \
  29             simgrid::instr::Container::get_root()->type_->by_name_or_create<simgrid::instr::EventType>(                \
  30                 _XBT_STRINGIFY(cat));                                                                                  \
  31                                                                                                                        \
  32         std::string cont_name = std::string("rank-" + std::to_string(simgrid::s4u::this_actor::get_pid()));            \
  33         type->add_entity_value(desc->name, "1.0 1.0 1.0");                                                             \
  34         new simgrid::instr::NewEvent(SIMIX_get_clock(), simgrid::instr::Container::by_name(cont_name), type,           \
  35                                      type->get_entity_value(desc->name));                                              \
  36       }                                                                                                                \
  37       time1 = SIMIX_get_clock();                                                                                       \
  38       try {                                                                                                            \
  39         ((int(*) args)desc->coll) args2;                                                                               \
  40       } catch (std::exception & ex) {                                                                                  \
  41         continue;                                                                                                      \
  42       }                                                                                                                \
  43       time2   = SIMIX_get_clock();                                                                                     \
  44       buf_out = time2 - time1;                                                                                         \
  45       reduce__default((void*)&buf_out, (void*)&buf_in, 1, MPI_DOUBLE, MPI_MAX, 0, comm);                               \
  46       if (time2 - time1 < time_min) {                                                                                  \
  47         min_coll = i;                                                                                                  \
  48         time_min = time2 - time1;                                                                                      \
  49       }                                                                                                                \
  50       if (comm->rank() == 0) {                                                                                         \
  51         if (buf_in < max_min) {                                                                                        \
  52           max_min     = buf_in;                                                                                        \
  53           global_coll = i;                                                                                             \
  54         }                                                                                                              \
  55       }                                                                                                                \
  56     next_iteration:                                                                                                    \
  57       i++;                                                                                                             \
  58       desc = simgrid::smpi::colls::get_smpi_coll_description(_XBT_STRINGIFY(cat), i);                                  \
  59     }                                                                                                                  \
  60     if (comm->rank() == 0) {                                                                                           \
  61       XBT_WARN("For rank 0, the quickest was %s : %f , but global was %s : %f at max",                                 \
  62                simgrid::smpi::colls::get_smpi_coll_description(_XBT_STRINGIFY(cat), min_coll)->name.c_str(), time_min, \
  63                simgrid::smpi::colls::get_smpi_coll_description(_XBT_STRINGIFY(cat), global_coll)->name.c_str(),        \
  64                max_min);                                                                                               \
  65     } else                                                                                                             \
  66       XBT_WARN("The quickest " _XBT_STRINGIFY(cat) " was %s on rank %d and took %f",                                   \
  67                simgrid::smpi::colls::get_smpi_coll_description(_XBT_STRINGIFY(cat), min_coll)->name.c_str(),           \
  68                comm->rank(), time_min);                                                                                \
  69     return (min_coll != -1) ? MPI_SUCCESS : MPI_ERR_INTERN;                                                            \
  70   }
  71
  72 namespace simgrid{
  73 namespace smpi{
  74
  75 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLGATHERV_SIG, (send_buff, send_count, send_type, recv_buff, recv_count, recv_disps, recv_type, comm))
  76 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLREDUCE_SIG, (sbuf, rbuf, rcount, dtype, op, comm))
  77 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_GATHER_SIG, (send_buff, send_count, send_type, recv_buff, recv_count, recv_type, root, comm))
  78 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLGATHER_SIG, (send_buff,send_count,send_type,recv_buff,recv_count,recv_type,comm))
  79 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_ALLTOALL_SIG,(send_buff, send_count, send_type, recv_buff, recv_count, recv_type,comm))
  80 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))
  81 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_BCAST_SIG , (buf, count, datatype, root, comm))
  82 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_REDUCE_SIG,(buf,rbuf, count, datatype, op, root, comm))
  83 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_REDUCE_SCATTER_SIG ,(sbuf,rbuf, rcounts,dtype,op,comm))
  84 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_SCATTER_SIG ,(sendbuf, sendcount, sendtype,recvbuf, recvcount, recvtype,root, comm))
  85 COLL_APPLY(AUTOMATIC_COLL_BENCH, COLL_BARRIER_SIG,(comm))
  86
  87 }
  88 }