src/smpi/colls/allgatherv-ompi-bruck.c

   1 /*
   2  * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
   3  *                         University Research and Technology
   4  *                         Corporation.  All rights reserved.
   5  * Copyright (c) 2004-2009 The University of Tennessee and The University
   6  *                         of Tennessee Research Foundation.  All rights
   7  *                         reserved.
   8  * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
   9  *                         University of Stuttgart.  All rights reserved.
  10  * Copyright (c) 2004-2005 The Regents of the University of California.
  11  *                         All rights reserved.
  12  * Copyright (c) 2009      University of Houston. All rights reserved.
  13  * $COPYRIGHT$
  14  *
  15  * Additional copyrights may follow
  16  *
  17  * $HEADER$
  18  */
  19
  20 #include "colls_private.h"
  21 #define  MCA_COLL_BASE_TAG_ALLGATHERV 444
  22 /*
  23  * ompi_coll_tuned_allgatherv_intra_bruck
  24  *
  25  * Function:     allgather using O(log(N)) steps.
  26  * Accepts:      Same arguments as MPI_Allgather
  27  * Returns:      MPI_SUCCESS or error code
  28  *
  29  * Description:  Variation to All-to-all algorithm described by Bruck et al.in
  30  *               "Efficient Algorithms for All-to-all Communications
  31  *                in Multiport Message-Passing Systems"
  32  * Note:         Unlike in case of allgather implementation, we relay on
  33  *               indexed datatype to select buffers appropriately.
  34  *               The only additional memory requirement is for creation of
  35  *               temporary datatypes.
  36  * Example on 7 nodes (memory lay out need not be in-order)
  37  *   Initial set up:
  38  *    #     0      1      2      3      4      5      6
  39  *         [0]    [ ]    [ ]    [ ]    [ ]    [ ]    [ ]
  40  *         [ ]    [1]    [ ]    [ ]    [ ]    [ ]    [ ]
  41  *         [ ]    [ ]    [2]    [ ]    [ ]    [ ]    [ ]
  42  *         [ ]    [ ]    [ ]    [3]    [ ]    [ ]    [ ]
  43  *         [ ]    [ ]    [ ]    [ ]    [4]    [ ]    [ ]
  44  *         [ ]    [ ]    [ ]    [ ]    [ ]    [5]    [ ]
  45  *         [ ]    [ ]    [ ]    [ ]    [ ]    [ ]    [6]
  46  *   Step 0: send message to (rank - 2^0), receive message from (rank + 2^0)
  47  *    #     0      1      2      3      4      5      6
  48  *         [0]    [ ]    [ ]    [ ]    [ ]    [ ]    [0]
  49  *         [1]    [1]    [ ]    [ ]    [ ]    [ ]    [ ]
  50  *         [ ]    [2]    [2]    [ ]    [ ]    [ ]    [ ]
  51  *         [ ]    [ ]    [3]    [3]    [ ]    [ ]    [ ]
  52  *         [ ]    [ ]    [ ]    [4]    [4]    [ ]    [ ]
  53  *         [ ]    [ ]    [ ]    [ ]    [5]    [5]    [ ]
  54  *         [ ]    [ ]    [ ]    [ ]    [ ]    [6]    [6]
  55  *   Step 1: send message to (rank - 2^1), receive message from (rank + 2^1).
  56  *           message contains all blocks from (rank) .. (rank + 2^2) with
  57  *           wrap around.
  58  *    #     0      1      2      3      4      5      6
  59  *         [0]    [ ]    [ ]    [ ]    [0]    [0]    [0]
  60  *         [1]    [1]    [ ]    [ ]    [ ]    [1]    [1]
  61  *         [2]    [2]    [2]    [ ]    [ ]    [ ]    [2]
  62  *         [3]    [3]    [3]    [3]    [ ]    [ ]    [ ]
  63  *         [ ]    [4]    [4]    [4]    [4]    [ ]    [ ]
  64  *         [ ]    [ ]    [5]    [5]    [5]    [5]    [ ]
  65  *         [ ]    [ ]    [ ]    [6]    [6]    [6]    [6]
  66  *   Step 2: send message to (rank - 2^2), receive message from (rank + 2^2).
  67  *           message size is "all remaining blocks"
  68  *    #     0      1      2      3      4      5      6
  69  *         [0]    [0]    [0]    [0]    [0]    [0]    [0]
  70  *         [1]    [1]    [1]    [1]    [1]    [1]    [1]
  71  *         [2]    [2]    [2]    [2]    [2]    [2]    [2]
  72  *         [3]    [3]    [3]    [3]    [3]    [3]    [3]
  73  *         [4]    [4]    [4]    [4]    [4]    [4]    [4]
  74  *         [5]    [5]    [5]    [5]    [5]    [5]    [5]
  75  *         [6]    [6]    [6]    [6]    [6]    [6]    [6]
  76  */
  77 int smpi_coll_tuned_allgatherv_ompi_bruck(void *sbuf, int scount,
  78                                            MPI_Datatype sdtype,
  79                                            void *rbuf, int *rcounts,
  80                                            int *rdispls,
  81                                            MPI_Datatype rdtype,
  82                                            MPI_Comm comm)
  83 {
  84    int rank, size;
  85    int sendto, recvfrom, distance, blockcount, i;
  86    int *new_rcounts = NULL, *new_rdispls = NULL;
  87    int *new_scounts = NULL, *new_sdispls = NULL;
  88    ptrdiff_t slb, rlb, sext, rext;
  89    char *tmpsend = NULL, *tmprecv = NULL;
  90    MPI_Datatype new_rdtype, new_sdtype;
  91
  92    size = smpi_comm_size(comm);
  93    rank = smpi_comm_rank(comm);
  94
  95    XBT_DEBUG(
  96                 "coll:tuned:allgather_ompi_bruck rank %d", rank);
  97
  98    smpi_datatype_extent (sdtype, &slb, &sext);
  99
 100    smpi_datatype_extent (rdtype, &rlb, &rext);
 101
 102    /* Initialization step:
 103       - if send buffer is not MPI_IN_PLACE, copy send buffer to block rank of
 104         the receive buffer.
 105    */
 106    tmprecv = (char*) rbuf + rdispls[rank] * rext;
 107    if (MPI_IN_PLACE != sbuf) {
 108       tmpsend = (char*) sbuf;
 109       smpi_datatype_copy(tmpsend, scount, sdtype,
 110                             tmprecv, rcounts[rank], rdtype);
 111    }
 112
 113    /* Communication step:
 114       At every step i, rank r:
 115       - doubles the distance
 116       - sends message with blockcount blocks, (rbuf[rank] .. rbuf[rank + 2^i])
 117         to rank (r - distance)
 118       - receives message of blockcount blocks,
 119         (rbuf[r + distance] ... rbuf[(r+distance) + 2^i]) from
 120         rank (r + distance)
 121       - blockcount doubles until the last step when only the remaining data is
 122       exchanged.
 123    */
 124    blockcount = 1;
 125    tmpsend = (char*) rbuf;
 126
 127    new_rcounts = (int*) calloc(4*size, sizeof(int));
 128    new_rdispls = new_rcounts + size;
 129    new_scounts = new_rdispls + size;
 130    new_sdispls = new_scounts + size;
 131
 132    for (distance = 1; distance < size; distance<<=1) {
 133
 134       recvfrom = (rank + distance) % size;
 135       sendto = (rank - distance + size) % size;
 136
 137       if (distance <= (size >> 1)) {
 138          blockcount = distance;
 139       } else {
 140          blockcount = size - distance;
 141       }
 142
 143       /* create send and receive datatypes */
 144       for (i = 0; i < blockcount; i++) {
 145           const int tmp_srank = (rank + i) % size;
 146           const int tmp_rrank = (recvfrom + i) % size;
 147           new_scounts[i] = rcounts[tmp_srank];
 148           new_sdispls[i] = rdispls[tmp_srank];
 149           new_rcounts[i] = rcounts[tmp_rrank];
 150           new_rdispls[i] = rdispls[tmp_rrank];
 151       }
 152       smpi_datatype_indexed(blockcount, new_scounts, new_sdispls,
 153                                     rdtype, &new_sdtype);
 154       smpi_datatype_indexed(blockcount, new_rcounts, new_rdispls,
 155                                     rdtype, &new_rdtype);
 156
 157       smpi_datatype_commit(&new_sdtype);
 158       smpi_datatype_commit(&new_rdtype);
 159
 160       /* Sendreceive */
 161       smpi_mpi_sendrecv(rbuf, 1, new_sdtype, sendto,
 162                                      MCA_COLL_BASE_TAG_ALLGATHERV,
 163                                      rbuf, 1, new_rdtype, recvfrom,
 164                                      MCA_COLL_BASE_TAG_ALLGATHERV,
 165                                      comm, MPI_STATUS_IGNORE);
 166       smpi_datatype_free(&new_sdtype);
 167       smpi_datatype_free(&new_rdtype);
 168
 169    }
 170
 171    free(new_rcounts);
 172
 173    return MPI_SUCCESS;
 174
 175 }
 176