src/smpi/colls/smpi_automatic_selector.c

   1 /* Copyright (c) 2013-2014. 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 "colls_private.h"
   8 #ifdef HAVE_MC
   9 #include "mc/mc_private.h"
  10 #endif
  11 #include <float.h>
  12
  13 //attempt to do a quick autotuning version of the collective,
  14
  15 #ifdef HAVE_TRACING
  16 #define TRACE_AUTO_COLL(cat) if (TRACE_is_enabled()){\
  17         type_t type = PJ_type_get_or_null (#cat, PJ_type_get_root());\
  18          if (!type){\
  19              type=PJ_type_event_new(#cat, PJ_type_get_root());\
  20          }\
  21          char cont_name[25];\
  22          sprintf(cont_name, "rank-%d", smpi_process_index());\
  23          val_t value = PJ_value_get_or_new(mpi_coll_##cat##_description[i].name,"1.0 1.0 1.0", type);\
  24          new_pajeNewEvent (SIMIX_get_clock(), PJ_container_get(cont_name), type, value);\
  25       }
  26 #else
  27 #define TRACE_AUTO_COLL(cat)
  28 #endif
  29
  30
  31 #define AUTOMATIC_COLL_BENCH(cat, ret, args, args2)\
  32     ret smpi_coll_tuned_ ## cat ## _ ## automatic(COLL_UNPAREN args)\
  33 {\
  34   double time1, time2, time_min=DBL_MAX;\
  35   volatile int min_coll=-1, global_coll=-1;\
  36   volatile int i;\
  37   xbt_ex_t ex;\
  38   double buf_in, buf_out, max_min=DBL_MAX;\
  39   for (i = 0; mpi_coll_##cat##_description[i].name; i++){\
  40       if(!strcmp(mpi_coll_##cat##_description[i].name, "automatic"))continue;\
  41       if(!strcmp(mpi_coll_##cat##_description[i].name, "default"))continue;\
  42       smpi_mpi_barrier(comm);\
  43       TRACE_AUTO_COLL(cat)\
  44       time1 = SIMIX_get_clock();\
  45       TRY{\
  46       ((int (*) args)\
  47           mpi_coll_##cat##_description[i].coll) args2 ;\
  48       }\
  49       CATCH(ex) {\
  50         xbt_ex_free(ex);\
  51         continue;\
  52       }\
  53       time2 = SIMIX_get_clock();\
  54       buf_out=time2-time1;\
  55       smpi_mpi_reduce((void*)&buf_out,(void*)&buf_in, 1, MPI_DOUBLE, MPI_MAX, 0,comm );\
  56       if(time2-time1<time_min){\
  57           min_coll=i;\
  58           time_min=time2-time1;\
  59       }\
  60       if(smpi_comm_rank(comm)==0){\
  61           if(buf_in<max_min){\
  62               max_min=buf_in;\
  63               global_coll=i;\
  64           }\
  65       }\
  66   }\
  67   if(smpi_comm_rank(comm)==0){\
  68       XBT_WARN("For rank 0, the quickest was %s : %f , but global was %s : %f at max",mpi_coll_##cat##_description[min_coll].name, time_min,mpi_coll_##cat##_description[global_coll].name, max_min);\
  69   }else\
  70   XBT_WARN("The quickest %s was %s on rank %d and took %f",#cat,mpi_coll_##cat##_description[min_coll].name, smpi_comm_rank(comm), time_min);\
  71   return (min_coll!=-1)?MPI_SUCCESS:MPI_ERR_INTERN;\
  72 }\
  73
  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));