1 /* Copyright (c) 2013-2021. 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) 2004-2006 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.
20 * Additional copyrights may follow
23 #include "../coll_tuned_topo.hpp"
24 #include "../colls_private.hpp"
27 * Barrier is meant to be a synchronous operation, as some BTLs can mark
28 * a request done before its passed to the NIC and progress might not be made
29 * elsewhere we cannot allow a process to exit the barrier until its last
30 * [round of] sends are completed.
32 * It is last round of sends rather than 'last' individual send as each pair of
33 * peers can use different channels/devices/btls and the receiver of one of
34 * these sends might be forced to wait as the sender
35 * leaves the collective and does not make progress until the next mpi call
40 * Simple double ring version of barrier
42 * synchronous guarantee made by last ring of sends are synchronous
47 int barrier__ompi_doublering(MPI_Comm comm)
56 XBT_DEBUG("ompi_coll_tuned_barrier_ompi_doublering rank %d", rank);
58 left = ((rank-1+size)%size);
59 right = ((rank+1)%size);
61 if (rank > 0) { /* receive message from the left */
62 Request::recv(nullptr, 0, MPI_BYTE, left, COLL_TAG_BARRIER, comm, MPI_STATUS_IGNORE);
65 /* Send message to the right */
66 Request::send(nullptr, 0, MPI_BYTE, right, COLL_TAG_BARRIER, comm);
68 /* root needs to receive from the last node */
70 Request::recv(nullptr, 0, MPI_BYTE, left, COLL_TAG_BARRIER, comm, MPI_STATUS_IGNORE);
73 /* Allow nodes to exit */
74 if (rank > 0) { /* post Receive from left */
75 Request::recv(nullptr, 0, MPI_BYTE, left, COLL_TAG_BARRIER, comm, MPI_STATUS_IGNORE);
78 /* send message to the right one */
79 Request::send(nullptr, 0, MPI_BYTE, right, COLL_TAG_BARRIER, comm);
81 /* rank 0 post receive from the last node */
83 Request::recv(nullptr, 0, MPI_BYTE, left, COLL_TAG_BARRIER, comm, MPI_STATUS_IGNORE);
92 * To make synchronous, uses sync sends and sync sendrecvs
95 int barrier__ompi_recursivedoubling(MPI_Comm comm)
97 int rank, size, adjsize;
103 "ompi_coll_tuned_barrier_ompi_recursivedoubling rank %d",
106 /* do nearest power of 2 less than size calc */
107 for( adjsize = 1; adjsize <= size; adjsize <<= 1 );
110 /* if size is not exact power of two, perform an extra step */
111 if (adjsize != size) {
112 if (rank >= adjsize) {
113 /* send message to lower ranked node */
114 remote = rank - adjsize;
115 Request::sendrecv(nullptr, 0, MPI_BYTE, remote, COLL_TAG_BARRIER, nullptr, 0, MPI_BYTE, remote,
116 COLL_TAG_BARRIER, comm, MPI_STATUS_IGNORE);
118 } else if (rank < (size - adjsize)) {
120 /* receive message from high level rank */
121 Request::recv(nullptr, 0, MPI_BYTE, rank + adjsize, COLL_TAG_BARRIER, comm, MPI_STATUS_IGNORE);
125 /* exchange messages */
126 if ( rank < adjsize ) {
128 while ( mask < adjsize ) {
129 remote = rank ^ mask;
131 if (remote >= adjsize) continue;
133 /* post receive from the remote node */
134 Request::sendrecv(nullptr, 0, MPI_BYTE, remote, COLL_TAG_BARRIER, nullptr, 0, MPI_BYTE, remote,
135 COLL_TAG_BARRIER, comm, MPI_STATUS_IGNORE);
139 /* non-power of 2 case */
140 if (adjsize != size) {
141 if (rank < (size - adjsize)) {
142 /* send enter message to higher ranked node */
143 remote = rank + adjsize;
144 Request::send(nullptr, 0, MPI_BYTE, remote, COLL_TAG_BARRIER, comm);
154 * To make synchronous, uses sync sends and sync sendrecvs
157 int barrier__ompi_bruck(MPI_Comm comm)
160 int distance, to, from;
165 "ompi_coll_tuned_barrier_ompi_bruck rank %d", rank);
167 /* exchange data with rank-2^k and rank+2^k */
168 for (distance = 1; distance < size; distance <<= 1) {
169 from = (rank + size - distance) % size;
170 to = (rank + distance) % size;
172 /* send message to lower ranked node */
173 Request::sendrecv(nullptr, 0, MPI_BYTE, to, COLL_TAG_BARRIER, nullptr, 0, MPI_BYTE, from, COLL_TAG_BARRIER,
174 comm, MPI_STATUS_IGNORE);
183 * To make synchronous, uses sync sends and sync sendrecvs
185 /* special case for two processes */
186 int barrier__ompi_two_procs(MPI_Comm comm)
190 remote = comm->rank();
192 "ompi_coll_tuned_barrier_ompi_two_procs rank %d", remote);
193 remote = (remote + 1) & 0x1;
195 Request::sendrecv(nullptr, 0, MPI_BYTE, remote, COLL_TAG_BARRIER, nullptr, 0, MPI_BYTE, remote, COLL_TAG_BARRIER,
196 comm, MPI_STATUS_IGNORE);
197 return (MPI_SUCCESS);
202 * Linear functions are copied from the BASIC coll module
203 * they do not segment the message and are simple implementations
204 * but for some small number of nodes and/or small data sizes they
205 * are just as fast as tuned/tree based segmenting operations
206 * and as such may be selected by the decision functions
207 * These are copied into this module due to the way we select modules
208 * in V1. i.e. in V2 we will handle this differently and so will not
209 * have to duplicate code.
210 * GEF Oct05 after asking Jeff.
213 /* copied function (with appropriate renaming) starts here */
215 int barrier__ompi_basic_linear(MPI_Comm comm)
218 int size = comm->size();
219 int rank = comm->rank();
221 /* All non-root send & receive zero-length message. */
224 Request::send(nullptr, 0, MPI_BYTE, 0, COLL_TAG_BARRIER, comm);
226 Request::recv(nullptr, 0, MPI_BYTE, 0, COLL_TAG_BARRIER, comm, MPI_STATUS_IGNORE);
229 /* The root collects and broadcasts the messages. */
232 MPI_Request* requests;
234 requests = new MPI_Request[size];
235 for (i = 1; i < size; ++i) {
236 requests[i] = Request::irecv(nullptr, 0, MPI_BYTE, i, COLL_TAG_BARRIER, comm);
238 Request::waitall( size-1, requests+1, MPI_STATUSES_IGNORE );
240 for (i = 1; i < size; ++i) {
241 requests[i] = Request::isend(nullptr, 0, MPI_BYTE, i, COLL_TAG_BARRIER, comm);
243 Request::waitall( size-1, requests+1, MPI_STATUSES_IGNORE );
252 /* copied function (with appropriate renaming) ends here */
255 * Another recursive doubling type algorithm, but in this case
256 * we go up the tree and back down the tree.
258 int barrier__ompi_tree(MPI_Comm comm)
260 int rank, size, depth;
266 "ompi_coll_tuned_barrier_ompi_tree %d",
269 /* Find the nearest power of 2 of the communicator size. */
270 for(depth = 1; depth < size; depth <<= 1 );
272 for (jump=1; jump<depth; jump<<=1) {
273 partner = rank ^ jump;
274 if (!(partner & (jump-1)) && partner < size) {
275 if (partner > rank) {
276 Request::recv(nullptr, 0, MPI_BYTE, partner, COLL_TAG_BARRIER, comm, MPI_STATUS_IGNORE);
277 } else if (partner < rank) {
278 Request::send(nullptr, 0, MPI_BYTE, partner, COLL_TAG_BARRIER, comm);
284 for (jump = depth; jump>0; jump>>=1) {
285 partner = rank ^ jump;
286 if (!(partner & (jump-1)) && partner < size) {
287 if (partner > rank) {
288 Request::send(nullptr, 0, MPI_BYTE, partner, COLL_TAG_BARRIER, comm);
289 } else if (partner < rank) {
290 Request::recv(nullptr, 0, MPI_BYTE, partner, COLL_TAG_BARRIER, comm, MPI_STATUS_IGNORE);