1 /* Copyright (c) 2013-2017. 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 ment 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 gurantee made by last ring of sends are synchronous
47 int Coll_barrier_ompi_doublering::barrier(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((void*)NULL, 0, MPI_BYTE, left,
63 COLL_TAG_BARRIER, comm,
67 /* Send message to the right */
68 Request::send((void*)NULL, 0, MPI_BYTE, right,
72 /* root needs to receive from the last node */
74 Request::recv((void*)NULL, 0, MPI_BYTE, left,
75 COLL_TAG_BARRIER, comm,
79 /* Allow nodes to exit */
80 if (rank > 0) { /* post Receive from left */
81 Request::recv((void*)NULL, 0, MPI_BYTE, left,
82 COLL_TAG_BARRIER, comm,
86 /* send message to the right one */
87 Request::send((void*)NULL, 0, MPI_BYTE, right,
91 /* rank 0 post receive from the last node */
93 Request::recv((void*)NULL, 0, MPI_BYTE, left,
94 COLL_TAG_BARRIER, comm,
104 * To make synchronous, uses sync sends and sync sendrecvs
107 int Coll_barrier_ompi_recursivedoubling::barrier(MPI_Comm comm)
109 int rank, size, adjsize;
115 "ompi_coll_tuned_barrier_ompi_recursivedoubling rank %d",
118 /* do nearest power of 2 less than size calc */
119 for( adjsize = 1; adjsize <= size; adjsize <<= 1 );
122 /* if size is not exact power of two, perform an extra step */
123 if (adjsize != size) {
124 if (rank >= adjsize) {
125 /* send message to lower ranked node */
126 remote = rank - adjsize;
127 Request::sendrecv(NULL, 0, MPI_BYTE, remote,
129 NULL, 0, MPI_BYTE, remote,
131 comm, MPI_STATUS_IGNORE);
133 } else if (rank < (size - adjsize)) {
135 /* receive message from high level rank */
136 Request::recv((void*)NULL, 0, MPI_BYTE, rank+adjsize,
137 COLL_TAG_BARRIER, comm,
143 /* exchange messages */
144 if ( rank < adjsize ) {
146 while ( mask < adjsize ) {
147 remote = rank ^ mask;
149 if (remote >= adjsize) continue;
151 /* post receive from the remote node */
152 Request::sendrecv(NULL, 0, MPI_BYTE, remote,
154 NULL, 0, MPI_BYTE, remote,
156 comm, MPI_STATUS_IGNORE);
160 /* non-power of 2 case */
161 if (adjsize != size) {
162 if (rank < (size - adjsize)) {
163 /* send enter message to higher ranked node */
164 remote = rank + adjsize;
165 Request::send((void*)NULL, 0, MPI_BYTE, remote,
178 * To make synchronous, uses sync sends and sync sendrecvs
181 int Coll_barrier_ompi_bruck::barrier(MPI_Comm comm)
184 int distance, to, from;
189 "ompi_coll_tuned_barrier_ompi_bruck rank %d", rank);
191 /* exchange data with rank-2^k and rank+2^k */
192 for (distance = 1; distance < size; distance <<= 1) {
193 from = (rank + size - distance) % size;
194 to = (rank + distance) % size;
196 /* send message to lower ranked node */
197 Request::sendrecv(NULL, 0, MPI_BYTE, to,
199 NULL, 0, MPI_BYTE, from,
201 comm, MPI_STATUS_IGNORE);
210 * To make synchronous, uses sync sends and sync sendrecvs
212 /* special case for two processes */
213 int Coll_barrier_ompi_two_procs::barrier(MPI_Comm comm)
217 remote = comm->rank();
219 "ompi_coll_tuned_barrier_ompi_two_procs rank %d", remote);
220 remote = (remote + 1) & 0x1;
222 Request::sendrecv(NULL, 0, MPI_BYTE, remote,
224 NULL, 0, MPI_BYTE, remote,
226 comm, MPI_STATUS_IGNORE);
227 return (MPI_SUCCESS);
232 * Linear functions are copied from the BASIC coll module
233 * they do not segment the message and are simple implementations
234 * but for some small number of nodes and/or small data sizes they
235 * are just as fast as tuned/tree based segmenting operations
236 * and as such may be selected by the decision functions
237 * These are copied into this module due to the way we select modules
238 * in V1. i.e. in V2 we will handle this differently and so will not
239 * have to duplicate code.
240 * GEF Oct05 after asking Jeff.
243 /* copied function (with appropriate renaming) starts here */
245 int Coll_barrier_ompi_basic_linear::barrier(MPI_Comm comm)
248 int size = comm->size();
249 int rank = comm->rank();
251 /* All non-root send & receive zero-length message. */
254 Request::send (NULL, 0, MPI_BYTE, 0,
258 Request::recv (NULL, 0, MPI_BYTE, 0,
260 comm, MPI_STATUS_IGNORE);
263 /* The root collects and broadcasts the messages. */
266 MPI_Request* requests;
268 requests = new MPI_Request[size];
269 for (i = 1; i < size; ++i) {
270 requests[i] = Request::irecv(NULL, 0, MPI_BYTE, MPI_ANY_SOURCE,
271 COLL_TAG_BARRIER, comm
274 Request::waitall( size-1, requests+1, MPI_STATUSES_IGNORE );
276 for (i = 1; i < size; ++i) {
277 requests[i] = Request::isend(NULL, 0, MPI_BYTE, i,
282 Request::waitall( size-1, requests+1, MPI_STATUSES_IGNORE );
291 /* copied function (with appropriate renaming) ends here */
294 * Another recursive doubling type algorithm, but in this case
295 * we go up the tree and back down the tree.
297 int Coll_barrier_ompi_tree::barrier(MPI_Comm comm)
299 int rank, size, depth;
305 "ompi_coll_tuned_barrier_ompi_tree %d",
308 /* Find the nearest power of 2 of the communicator size. */
309 for(depth = 1; depth < size; depth <<= 1 );
311 for (jump=1; jump<depth; jump<<=1) {
312 partner = rank ^ jump;
313 if (!(partner & (jump-1)) && partner < size) {
314 if (partner > rank) {
315 Request::recv (NULL, 0, MPI_BYTE, partner,
316 COLL_TAG_BARRIER, comm,
318 } else if (partner < rank) {
319 Request::send (NULL, 0, MPI_BYTE, partner,
327 for (jump = depth; jump>0; jump>>=1) {
328 partner = rank ^ jump;
329 if (!(partner & (jump-1)) && partner < size) {
330 if (partner > rank) {
331 Request::send (NULL, 0, MPI_BYTE, partner,
334 } else if (partner < rank) {
335 Request::recv (NULL, 0, MPI_BYTE, partner,
336 COLL_TAG_BARRIER, comm,