Logo AND Algorithmique Numérique Distribuée

Public GIT Repository
refactor with templated method
[simgrid.git] / src / smpi / colls / smpi_automatic_selector.cpp
1 /* Copyright (c) 2013-2018. 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 =                                                                                  \
17         simgrid::instr::Container::getRoot()->type_->by_name_or_create<simgrid::instr::EventType>(#cat);               \
18                                                                                                                        \
19     std::string cont_name = std::string("rank-" + std::to_string(simgrid::s4u::this_actor::get_pid()));                \
20     type->addEntityValue(Colls::mpi_coll_##cat##_description[i].name, "1.0 1.0 1.0");                                  \
21     new simgrid::instr::NewEvent(SIMIX_get_clock(), simgrid::instr::Container::byName(cont_name), type,                \
22                                  type->getEntityValue(Colls::mpi_coll_##cat##_description[i].name));                   \
23   }
24
25 #define AUTOMATIC_COLL_BENCH(cat, ret, args, args2)                                                                    \
26   ret Coll_##cat##_automatic::cat(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; Colls::mpi_coll_##cat##_description[i].name; i++) {                                                    \
33       if (not strcmp(Colls::mpi_coll_##cat##_description[i].name, "automatic"))                                        \
34         continue;                                                                                                      \
35       if (not strcmp(Colls::mpi_coll_##cat##_description[i].name, "default"))                                          \
36         continue;                                                                                                      \
37       Coll_barrier_default::barrier(comm);                                                                             \
38       TRACE_AUTO_COLL(cat)                                                                                             \
39       time1 = SIMIX_get_clock();                                                                                       \
40       try {                                                                                                            \
41         ((int(*) args)Colls::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       Coll_reduce_default::reduce((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::mpi_coll_##cat##_description[min_coll].name, time_min,                                           \
62                Colls::mpi_coll_##cat##_description[global_coll].name, max_min);                                        \
63     } else                                                                                                             \
64       XBT_WARN("The quickest %s was %s on rank %d and took %f", #cat,                                                  \
65                Colls::mpi_coll_##cat##_description[min_coll].name, 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 }