1 /* Copyright (c) 2013-2020. The SimGrid Team.
2 * All rights reserved. */
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. */
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) 2004-2012 The University of Tennessee and The University
12 * of Tennessee Research Foundation. All rights
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 * Copyright (c) 2008 Sun Microsystems, Inc. All rights reserved.
19 * Copyright (c) 2009 University of Houston. All rights reserved.
21 * Additional copyrights may follow
26 * (C) 2001 by Argonne National Laboratory.
27 * See COPYRIGHT in top-level directory.
29 /* Copyright (c) 2001-2014, The Ohio State University. All rights
32 * This file is part of the MVAPICH2 software package developed by the
33 * team members of The Ohio State University's Network-Based Computing
34 * Laboratory (NBCL), headed by Professor Dhabaleswar K. (DK) Panda.
36 * For detailed copyright and licensing information, please refer to the
37 * copyright file COPYRIGHT in the top level MVAPICH2 directory.
42 #define MV2_ALLTOALL_THROTTLE_FACTOR 4
44 #include "../colls_private.hpp"
47 int alltoall__mvapich2_scatter_dest(const void *sendbuf,
49 MPI_Datatype sendtype,
52 MPI_Datatype recvtype,
56 MPI_Aint sendtype_extent = 0, recvtype_extent = 0;
57 int mpi_errno=MPI_SUCCESS;
60 if (recvcount == 0) return MPI_SUCCESS;
62 comm_size = comm->size();
65 /* Get extent of send and recv types */
66 recvtype_extent = recvtype->get_extent();
67 sendtype_extent = sendtype->get_extent();
69 /* Medium-size message. Use isend/irecv with scattered
70 destinations. Use Tony Ladd's modification to post only
71 a small number of isends/irecvs at a time. */
72 /* FIXME: This converts the Alltoall to a set of blocking phases.
73 Two alternatives should be considered:
74 1) the choice of communication pattern could try to avoid
75 contending routes in each phase
76 2) rather than wait for all communication to finish (waitall),
77 we could maintain constant queue size by using waitsome
78 and posting new isend/irecv as others complete. This avoids
79 synchronization delays at the end of each block (when
80 there are only a few isend/irecvs left)
84 //Stampede is configured with
85 bblock = MV2_ALLTOALL_THROTTLE_FACTOR;//mv2_coll_param.alltoall_throttle_factor;
87 if (bblock >= comm_size) bblock = comm_size;
88 /* If throttle_factor is n, each process posts n pairs of isend/irecv
91 /* FIXME: This should use the memory macros (there are storage
92 leaks here if there is an error, for example) */
93 auto* reqarray = new MPI_Request[2 * bblock];
95 auto* starray = new MPI_Status[2 * bblock];
97 for (ii=0; ii<comm_size; ii+=bblock) {
98 ss = comm_size-ii < bblock ? comm_size-ii : bblock;
99 /* do the communication -- post ss sends and receives: */
100 for ( i=0; i<ss; i++ ) {
101 dst = (rank+i+ii) % comm_size;
102 reqarray[i]=Request::irecv((char *)recvbuf +
103 dst*recvcount*recvtype_extent,
104 recvcount, recvtype, dst,
105 COLL_TAG_ALLTOALL, comm);
108 for ( i=0; i<ss; i++ ) {
109 dst = (rank-i-ii+comm_size) % comm_size;
110 reqarray[i+ss]=Request::isend((char *)sendbuf +
111 dst*sendcount*sendtype_extent,
112 sendcount, sendtype, dst,
113 COLL_TAG_ALLTOALL, comm);
117 /* ... then wait for them to finish: */
118 Request::waitall(2*ss,reqarray,starray);
121 /* --BEGIN ERROR HANDLING-- */
122 if (mpi_errno == MPI_ERR_IN_STATUS) {
123 for (j=0; j<2*ss; j++) {
124 if (starray[j].MPI_ERROR != MPI_SUCCESS) {
125 mpi_errno = starray[j].MPI_ERROR;
130 /* --END ERROR HANDLING-- */