src/smpi/colls/bcast/bcast-mvapich-smp.cpp

   1 /* Copyright (c) 2013-2019. 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 /*
   8  * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
   9  *                         University Research and Technology
  10  *                         Corporation.  All rights reserved.
  11  * Copyright (c) 2004-2009 The University of Tennessee and The University
  12  *                         of Tennessee Research Foundation.  All rights
  13  *                         reserved.
  14  * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
  15  *                         University of Stuttgart.  All rights reserved.
  16  * Copyright (c) 2004-2005 The Regents of the University of California.
  17  *                         All rights reserved.
  18  *
  19  * Additional copyrights may follow
  20  */
  21  /* -*- Mode: C; c-basic-offset:4 ; -*- */
  22 /* Copyright (c) 2001-2014, The Ohio State University. All rights
  23  * reserved.
  24  *
  25  * This file is part of the MVAPICH2 software package developed by the
  26  * team members of The Ohio State University's Network-Based Computing
  27  * Laboratory (NBCL), headed by Professor Dhabaleswar K. (DK) Panda.
  28  *
  29  * For detailed copyright and licensing information, please refer to the
  30  * copyright file COPYRIGHT in the top level MVAPICH2 directory.
  31  */
  32 /*
  33  *
  34  *  (C) 2001 by Argonne National Laboratory.
  35  *      See COPYRIGHT in top-level directory.
  36  */
  37
  38 #include "../colls_private.hpp"
  39
  40 extern int (*MV2_Bcast_function) (void *buffer, int count, MPI_Datatype datatype,
  41                            int root, MPI_Comm comm_ptr);
  42
  43 extern int (*MV2_Bcast_intra_node_function) (void *buffer, int count, MPI_Datatype datatype,
  44                                       int root, MPI_Comm comm_ptr);
  45
  46 extern int zcpy_knomial_factor;
  47 extern int mv2_pipelined_zcpy_knomial_factor;
  48 extern int bcast_segment_size;
  49 extern int mv2_inter_node_knomial_factor;
  50 extern int mv2_intra_node_knomial_factor;
  51 extern int mv2_bcast_two_level_system_size;
  52 #define INTRA_NODE_ROOT 0
  53
  54 #define MPIR_Pipelined_Bcast_Zcpy_MV2 Coll_bcast_mpich::bcast
  55 #define MPIR_Pipelined_Bcast_MV2 Coll_bcast_mpich::bcast
  56 #define MPIR_Bcast_binomial_MV2 Coll_bcast_binomial_tree::bcast
  57 #define MPIR_Bcast_scatter_ring_allgather_shm_MV2 Coll_bcast_scatter_LR_allgather::bcast
  58 #define MPIR_Bcast_scatter_doubling_allgather_MV2 Coll_bcast_scatter_rdb_allgather::bcast
  59 #define MPIR_Bcast_scatter_ring_allgather_MV2 Coll_bcast_scatter_LR_allgather::bcast
  60 #define MPIR_Shmem_Bcast_MV2 Coll_bcast_mpich::bcast
  61 #define MPIR_Bcast_tune_inter_node_helper_MV2 Coll_bcast_mvapich2_inter_node::bcast
  62 #define MPIR_Bcast_inter_node_helper_MV2 Coll_bcast_mvapich2_inter_node::bcast
  63 #define MPIR_Knomial_Bcast_intra_node_MV2 Coll_bcast_mvapich2_knomial_intra_node::bcast
  64 #define MPIR_Bcast_intra_MV2 Coll_bcast_mvapich2_intra_node::bcast
  65
  66 extern int zcpy_knomial_factor;
  67 extern int mv2_pipelined_zcpy_knomial_factor;
  68 extern int bcast_segment_size;
  69 extern int mv2_inter_node_knomial_factor;
  70 extern int mv2_intra_node_knomial_factor;
  71 #define mv2_bcast_two_level_system_size  64
  72 #define mv2_bcast_short_msg             16384
  73 #define mv2_bcast_large_msg            512*1024
  74 #define mv2_knomial_intra_node_threshold 131072
  75 #define mv2_scatter_rd_inter_leader_bcast 1
  76 namespace simgrid{
  77 namespace smpi{
  78 int Coll_bcast_mvapich2_inter_node::bcast(void *buffer,
  79                                                  int count,
  80                                                  MPI_Datatype datatype,
  81                                                  int root,
  82                                                  MPI_Comm  comm)
  83 {
  84     int rank;
  85     int mpi_errno = MPI_SUCCESS;
  86     MPI_Comm shmem_comm, leader_comm;
  87     int local_rank, local_size, global_rank = -1;
  88     int leader_root, leader_of_root;
  89
  90
  91     rank = comm->rank();
  92     //comm_size = comm->size();
  93
  94
  95     if (MV2_Bcast_function==NULL){
  96       MV2_Bcast_function=Coll_bcast_mpich::bcast;
  97     }
  98
  99     if (MV2_Bcast_intra_node_function==NULL){
 100       MV2_Bcast_intra_node_function= Coll_bcast_mpich::bcast;
 101     }
 102
 103     if(comm->get_leaders_comm()==MPI_COMM_NULL){
 104       comm->init_smp();
 105     }
 106
 107     shmem_comm = comm->get_intra_comm();
 108     local_rank = shmem_comm->rank();
 109     local_size = shmem_comm->size();
 110
 111     leader_comm = comm->get_leaders_comm();
 112
 113     if ((local_rank == 0) && (local_size > 1)) {
 114       global_rank = leader_comm->rank();
 115     }
 116
 117     int* leaders_map = comm->get_leaders_map();
 118     leader_of_root = comm->group()->rank(leaders_map[root]);
 119     leader_root = leader_comm->group()->rank(leaders_map[root]);
 120
 121
 122     if (local_size > 1) {
 123         if ((local_rank == 0) && (root != rank) && (leader_root == global_rank)) {
 124             Request::recv(buffer, count, datatype, root,
 125                                      COLL_TAG_BCAST, comm, MPI_STATUS_IGNORE);
 126         }
 127         if ((local_rank != 0) && (root == rank)) {
 128             Request::send(buffer, count, datatype,
 129                                      leader_of_root, COLL_TAG_BCAST, comm);
 130         }
 131     }
 132 #if defined(_MCST_SUPPORT_)
 133     if (comm_ptr->ch.is_mcast_ok) {
 134         mpi_errno = MPIR_Mcast_inter_node_MV2(buffer, count, datatype, root, comm_ptr,
 135                                               errflag);
 136         if (mpi_errno == MPI_SUCCESS) {
 137             goto fn_exit;
 138         } else {
 139             goto fn_fail;
 140         }
 141     }
 142 #endif
 143 /*
 144     if (local_rank == 0) {
 145         leader_comm = comm->get_leaders_comm();
 146         root = leader_root;
 147     }
 148
 149     if (MV2_Bcast_function == &MPIR_Pipelined_Bcast_MV2) {
 150         mpi_errno = MPIR_Pipelined_Bcast_MV2(buffer, count, datatype,
 151                                              root, comm);
 152     } else if (MV2_Bcast_function == &MPIR_Bcast_scatter_ring_allgather_shm_MV2) {
 153         mpi_errno = MPIR_Bcast_scatter_ring_allgather_shm_MV2(buffer, count,
 154                                                               datatype, root,
 155                                                               comm);
 156     } else */{
 157         if (local_rank == 0) {
 158       /*      if (MV2_Bcast_function == &MPIR_Knomial_Bcast_inter_node_wrapper_MV2) {
 159                 mpi_errno = MPIR_Knomial_Bcast_inter_node_wrapper_MV2(buffer, count,
 160                                                               datatype, root,
 161                                                               comm);
 162             } else {*/
 163                 mpi_errno = MV2_Bcast_function(buffer, count, datatype,
 164                                                leader_root, leader_comm);
 165           //  }
 166         }
 167     }
 168
 169     return mpi_errno;
 170 }
 171
 172
 173 int Coll_bcast_mvapich2_knomial_intra_node::bcast(void *buffer,
 174                                       int count,
 175                                       MPI_Datatype datatype,
 176                                       int root, MPI_Comm  comm)
 177 {
 178     int local_size = 0, rank;
 179     int mpi_errno = MPI_SUCCESS;
 180     int src, dst, mask, relative_rank;
 181     int k;
 182     if (MV2_Bcast_function==NULL){
 183       MV2_Bcast_function=Coll_bcast_mpich::bcast;
 184     }
 185
 186     if (MV2_Bcast_intra_node_function==NULL){
 187       MV2_Bcast_intra_node_function= Coll_bcast_mpich::bcast;
 188     }
 189
 190     if(comm->get_leaders_comm()==MPI_COMM_NULL){
 191       comm->init_smp();
 192     }
 193
 194     local_size = comm->size();
 195     rank = comm->rank();
 196
 197     MPI_Request* reqarray = new MPI_Request[2 * mv2_intra_node_knomial_factor];
 198
 199     MPI_Status* starray = new MPI_Status[2 * mv2_intra_node_knomial_factor];
 200
 201     /* intra-node k-nomial bcast  */
 202     if (local_size > 1) {
 203         relative_rank = (rank >= root) ? rank - root : rank - root + local_size;
 204         mask = 0x1;
 205
 206         while (mask < local_size) {
 207             if (relative_rank % (mv2_intra_node_knomial_factor * mask)) {
 208                 src = relative_rank / (mv2_intra_node_knomial_factor * mask) *
 209                     (mv2_intra_node_knomial_factor * mask) + root;
 210                 if (src >= local_size) {
 211                     src -= local_size;
 212                 }
 213
 214                 Request::recv(buffer, count, datatype, src,
 215                                          COLL_TAG_BCAST, comm,
 216                                          MPI_STATUS_IGNORE);
 217                 break;
 218             }
 219             mask *= mv2_intra_node_knomial_factor;
 220         }
 221         mask /= mv2_intra_node_knomial_factor;
 222
 223         while (mask > 0) {
 224             int reqs = 0;
 225             for (k = 1; k < mv2_intra_node_knomial_factor; k++) {
 226                 if (relative_rank + mask * k < local_size) {
 227                     dst = rank + mask * k;
 228                     if (dst >= local_size) {
 229                         dst -= local_size;
 230                     }
 231                     reqarray[reqs++]=Request::isend(buffer, count, datatype, dst,
 232                                               COLL_TAG_BCAST, comm);
 233                 }
 234             }
 235             Request::waitall(reqs, reqarray, starray);
 236
 237             mask /= mv2_intra_node_knomial_factor;
 238         }
 239     }
 240     delete[] reqarray;
 241     delete[] starray;
 242     return mpi_errno;
 243 }
 244
 245
 246 int Coll_bcast_mvapich2_intra_node::bcast(void *buffer,
 247                          int count,
 248                          MPI_Datatype datatype,
 249                          int root, MPI_Comm  comm)
 250 {
 251     int mpi_errno = MPI_SUCCESS;
 252     int comm_size;
 253     int two_level_bcast = 1;
 254     size_t nbytes = 0;
 255     int is_homogeneous, is_contig;
 256     MPI_Aint type_size;
 257     void *tmp_buf = NULL;
 258     MPI_Comm shmem_comm;
 259
 260     if (count == 0)
 261         return MPI_SUCCESS;
 262     if (MV2_Bcast_function==NULL){
 263       MV2_Bcast_function=Coll_bcast_mpich::bcast;
 264     }
 265
 266     if (MV2_Bcast_intra_node_function==NULL){
 267       MV2_Bcast_intra_node_function= Coll_bcast_mpich::bcast;
 268     }
 269
 270     if(comm->get_leaders_comm()==MPI_COMM_NULL){
 271       comm->init_smp();
 272     }
 273
 274     comm_size = comm->size();
 275    // rank = comm->rank();
 276 /*
 277     if (HANDLE_GET_KIND(datatype) == HANDLE_KIND_BUILTIN)*/
 278         is_contig = 1;
 279 /*    else {
 280         MPID_Datatype_get_ptr(datatype, dtp);
 281         is_contig = dtp->is_contig;
 282     }
 283 */
 284     is_homogeneous = 1;
 285 #ifdef MPID_HAS_HETERO
 286     if (comm_ptr->is_hetero)
 287         is_homogeneous = 0;
 288 #endif
 289
 290     /* MPI_Type_size() might not give the accurate size of the packed
 291      * datatype for heterogeneous systems (because of padding, encoding,
 292      * etc). On the other hand, MPI_Pack_size() can become very
 293      * expensive, depending on the implementation, especially for
 294      * heterogeneous systems. We want to use MPI_Type_size() wherever
 295      * possible, and MPI_Pack_size() in other places.
 296      */
 297     //if (is_homogeneous) {
 298         type_size=datatype->size();
 299     //}
 300 /*    else {*/
 301 /*        MPIR_Pack_size_impl(1, datatype, &type_size);*/
 302 /*    }*/
 303     nbytes = (size_t) (count) * (type_size);
 304     if (comm_size <= mv2_bcast_two_level_system_size) {
 305         if (nbytes > mv2_bcast_short_msg && nbytes < mv2_bcast_large_msg) {
 306             two_level_bcast = 1;
 307         } else {
 308             two_level_bcast = 0;
 309         }
 310     }
 311
 312     if (two_level_bcast == 1
 313 #if defined(_MCST_SUPPORT_)
 314             || comm_ptr->ch.is_mcast_ok
 315 #endif
 316         ) {
 317
 318       if (not is_contig || not is_homogeneous) {
 319         tmp_buf = (void*)smpi_get_tmp_sendbuffer(nbytes);
 320
 321         /* TODO: Pipeline the packing and communication */
 322         // position = 0;
 323         /*            if (rank == root) {*/
 324         /*                mpi_errno =*/
 325         /*                    MPIR_Pack_impl(buffer, count, datatype, tmp_buf, nbytes, &position);*/
 326         /*                if (mpi_errno)*/
 327         /*                    MPIU_ERR_POP(mpi_errno);*/
 328         /*            }*/
 329         }
 330
 331         shmem_comm = comm->get_intra_comm();
 332         if (not is_contig || not is_homogeneous) {
 333           mpi_errno = MPIR_Bcast_inter_node_helper_MV2(tmp_buf, nbytes, MPI_BYTE, root, comm);
 334         } else {
 335             mpi_errno =
 336                 MPIR_Bcast_inter_node_helper_MV2(buffer, count, datatype, root,
 337                                                  comm);
 338         }
 339
 340         /* We are now done with the inter-node phase */
 341             if (nbytes <= mv2_knomial_intra_node_threshold) {
 342               if (not is_contig || not is_homogeneous) {
 343                 mpi_errno = MPIR_Shmem_Bcast_MV2(tmp_buf, nbytes, MPI_BYTE, root, shmem_comm);
 344                 } else {
 345                     mpi_errno = MPIR_Shmem_Bcast_MV2(buffer, count, datatype,
 346                                                      root, shmem_comm);
 347                 }
 348             } else {
 349               if (not is_contig || not is_homogeneous) {
 350                 mpi_errno = MPIR_Knomial_Bcast_intra_node_MV2(tmp_buf, nbytes, MPI_BYTE, INTRA_NODE_ROOT, shmem_comm);
 351                 } else {
 352                     mpi_errno =
 353                         MPIR_Knomial_Bcast_intra_node_MV2(buffer, count,
 354                                                           datatype,
 355                                                           INTRA_NODE_ROOT,
 356                                                           shmem_comm);
 357                 }
 358             }
 359
 360     } else {
 361         if (nbytes <= mv2_bcast_short_msg) {
 362             mpi_errno = MPIR_Bcast_binomial_MV2(buffer, count, datatype, root,
 363                                                 comm);
 364         } else {
 365             if (mv2_scatter_rd_inter_leader_bcast) {
 366                 mpi_errno = MPIR_Bcast_scatter_ring_allgather_MV2(buffer, count,
 367                                                                   datatype,
 368                                                                   root,
 369                                                                   comm);
 370             } else {
 371                 mpi_errno =
 372                     MPIR_Bcast_scatter_doubling_allgather_MV2(buffer, count,
 373                                                               datatype, root,
 374                                                               comm);
 375             }
 376         }
 377     }
 378
 379
 380     return mpi_errno;
 381
 382 }
 383
 384 }
 385 }