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) 2004-2009 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.
19 * Additional copyrights may follow
22 #include "../coll_tuned_topo.hpp"
23 #include "../colls_private.hpp"
28 int smpi_coll_tuned_ompi_reduce_generic(const void* sendbuf, void* recvbuf, int original_count,
29 MPI_Datatype datatype, MPI_Op op,
30 int root, MPI_Comm comm,
31 ompi_coll_tree_t* tree, int count_by_segment,
32 int max_outstanding_reqs );
34 * This is a generic implementation of the reduce protocol. It used the tree
35 * provided as an argument and execute all operations using a segment of
36 * count times a datatype.
37 * For the last communication it will update the count in order to limit
38 * the number of datatype to the original count (original_count)
40 * Note that for non-commutative operations we cannot save memory copy
41 * for the first block: thus we must copy sendbuf to accumbuf on intermediate
42 * to keep the optimized loop happy.
44 int smpi_coll_tuned_ompi_reduce_generic(const void* sendbuf, void* recvbuf, int original_count,
45 MPI_Datatype datatype, MPI_Op op,
46 int root, MPI_Comm comm,
47 ompi_coll_tree_t* tree, int count_by_segment,
48 int max_outstanding_reqs )
50 unsigned char *inbuf[2] = {nullptr, nullptr}, *inbuf_free[2] = {nullptr, nullptr};
51 unsigned char *accumbuf = nullptr, *accumbuf_free = nullptr;
52 const unsigned char *local_op_buffer = nullptr, *sendtmpbuf = nullptr;
53 ptrdiff_t extent, lower_bound, segment_increment;
54 MPI_Request reqs[2] = {MPI_REQUEST_NULL, MPI_REQUEST_NULL};
55 int num_segments, line, ret, segindex, i, rank;
56 int recvcount, prevcount, inbi;
59 * Determine number of segments and number of elements
62 datatype->extent(&lower_bound, &extent);
63 num_segments = (original_count + count_by_segment - 1) / count_by_segment;
64 segment_increment = count_by_segment * extent;
66 sendtmpbuf = static_cast<const unsigned char*>(sendbuf);
67 if (sendbuf == MPI_IN_PLACE) {
68 sendtmpbuf = static_cast<const unsigned char*>(recvbuf);
71 XBT_DEBUG("coll:tuned:reduce_generic count %d, msg size %lu, segsize %lu, max_requests %d", original_count,
72 (unsigned long)(num_segments * segment_increment), (unsigned long)segment_increment,
73 max_outstanding_reqs);
77 /* non-leaf nodes - wait for children to send me data & forward up
79 if( tree->tree_nextsize > 0 ) {
80 ptrdiff_t true_extent, real_segment_size;
81 true_extent=datatype->get_extent();
83 /* handle non existent recv buffer (i.e. its NULL) and
84 protect the recv buffer on non-root nodes */
85 accumbuf = static_cast<unsigned char*>(recvbuf);
86 if (nullptr == accumbuf || root != rank) {
87 /* Allocate temporary accumulator buffer. */
88 accumbuf_free = smpi_get_tmp_sendbuffer(true_extent + (original_count - 1) * extent);
89 if (accumbuf_free == nullptr) {
94 accumbuf = accumbuf_free - lower_bound;
97 /* If this is a non-commutative operation we must copy
98 sendbuf to the accumbuf, in order to simplify the loops */
99 if ((op != MPI_OP_NULL && not op->is_commutative())) {
100 Datatype::copy(sendtmpbuf, original_count, datatype, accumbuf, original_count, datatype);
102 /* Allocate two buffers for incoming segments */
103 real_segment_size = true_extent + (count_by_segment - 1) * extent;
104 inbuf_free[0] = smpi_get_tmp_recvbuffer(real_segment_size);
105 if (inbuf_free[0] == nullptr) {
110 inbuf[0] = inbuf_free[0] - lower_bound;
111 /* if there is chance to overlap communication -
112 allocate second buffer */
113 if( (num_segments > 1) || (tree->tree_nextsize > 1) ) {
114 inbuf_free[1] = smpi_get_tmp_recvbuffer(real_segment_size);
115 if (inbuf_free[1] == nullptr) {
120 inbuf[1] = inbuf_free[1] - lower_bound;
123 /* reset input buffer index and receive count */
126 /* for each segment */
127 for( segindex = 0; segindex <= num_segments; segindex++ ) {
128 prevcount = recvcount;
129 /* recvcount - number of elements in current segment */
130 recvcount = count_by_segment;
131 if( segindex == (num_segments-1) )
132 recvcount = original_count - count_by_segment * segindex;
135 for( i = 0; i < tree->tree_nextsize; i++ ) {
137 * We try to overlap communication:
138 * either with next segment or with the next child
140 /* post irecv for current segindex on current child */
141 if( segindex < num_segments ) {
142 void* local_recvbuf = inbuf[inbi];
144 /* for the first step (1st child per segment) and
145 * commutative operations we might be able to irecv
146 * directly into the accumulate buffer so that we can
147 * reduce(op) this with our sendbuf in one step as
148 * ompi_op_reduce only has two buffer pointers,
149 * this avoids an extra memory copy.
151 * BUT if the operation is non-commutative or
152 * we are root and are USING MPI_IN_PLACE this is wrong!
154 if( (op==MPI_OP_NULL || op->is_commutative()) &&
155 !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
156 local_recvbuf = accumbuf + segindex * segment_increment;
160 reqs[inbi]=Request::irecv(local_recvbuf, recvcount, datatype,
162 COLL_TAG_REDUCE, comm
165 /* wait for previous req to complete, if any.
166 if there are no requests reqs[inbi ^1] will be
168 /* wait on data from last child for previous segment */
169 Request::waitall( 1, &reqs[inbi ^ 1],
170 MPI_STATUSES_IGNORE );
171 local_op_buffer = inbuf[inbi ^ 1];
173 /* our first operation is to combine our own [sendbuf] data
174 * with the data we recvd from down stream (but only
175 * the operation is commutative and if we are not root and
176 * not using MPI_IN_PLACE)
179 if( (op==MPI_OP_NULL || op->is_commutative())&&
180 !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
181 local_op_buffer = sendtmpbuf + segindex * segment_increment;
184 /* apply operation */
185 if(op!=MPI_OP_NULL) op->apply( local_op_buffer,
186 accumbuf + segindex * segment_increment,
187 &recvcount, datatype );
188 } else if ( segindex > 0 ) {
189 void* accumulator = accumbuf + (segindex-1) * segment_increment;
190 if( tree->tree_nextsize <= 1 ) {
191 if( (op==MPI_OP_NULL || op->is_commutative()) &&
192 !((MPI_IN_PLACE == sendbuf) && (rank == tree->tree_root)) ) {
193 local_op_buffer = sendtmpbuf + (segindex-1) * segment_increment;
196 if(op!=MPI_OP_NULL) op->apply( local_op_buffer, accumulator, &prevcount,
199 /* all reduced on available data this step (i) complete,
200 * pass to the next process unless you are the root.
202 if (rank != tree->tree_root) {
203 /* send combined/accumulated data to parent */
204 Request::send( accumulator, prevcount,
205 datatype, tree->tree_prev,
210 /* we stop when segindex = number of segments
211 (i.e. we do num_segment+1 steps for pipelining */
212 if (segindex == num_segments) break;
215 /* update input buffer index */
217 } /* end of for each child */
218 } /* end of for each segment */
221 smpi_free_tmp_buffer(inbuf_free[0]);
222 smpi_free_tmp_buffer(inbuf_free[1]);
223 smpi_free_tmp_buffer(accumbuf_free);
227 Depending on the value of max_outstanding_reqs and
228 the number of segments we have two options:
229 - send all segments using blocking send to the parent, or
230 - avoid overflooding the parent nodes by limiting the number of
231 outstanding requests to max_oustanding_reqs.
232 TODO/POSSIBLE IMPROVEMENT: If there is a way to determine the eager size
233 for the current communication, synchronization should be used only
234 when the message/segment size is smaller than the eager size.
238 /* If the number of segments is less than a maximum number of oustanding
239 requests or there is no limit on the maximum number of outstanding
240 requests, we send data to the parent using blocking send */
241 if ((0 == max_outstanding_reqs) ||
242 (num_segments <= max_outstanding_reqs)) {
245 while ( original_count > 0) {
246 if (original_count < count_by_segment) {
247 count_by_segment = original_count;
249 Request::send((char*)sendbuf +
250 segindex * segment_increment,
251 count_by_segment, datatype,
256 original_count -= count_by_segment;
260 /* Otherwise, introduce flow control:
261 - post max_outstanding_reqs non-blocking synchronous send,
262 - for remaining segments
263 - wait for a ssend to complete, and post the next one.
264 - wait for all outstanding sends to complete.
269 MPI_Request* sreq = new (std::nothrow) MPI_Request[max_outstanding_reqs];
270 if (NULL == sreq) { line = __LINE__; ret = -1; goto error_hndl; }
272 /* post first group of requests */
273 for (segindex = 0; segindex < max_outstanding_reqs; segindex++) {
274 sreq[segindex]=Request::isend((char*)sendbuf +
275 segindex * segment_increment,
276 count_by_segment, datatype,
280 original_count -= count_by_segment;
284 while ( original_count > 0 ) {
285 /* wait on a posted request to complete */
286 Request::wait(&sreq[creq], MPI_STATUS_IGNORE);
287 sreq[creq] = MPI_REQUEST_NULL;
289 if( original_count < count_by_segment ) {
290 count_by_segment = original_count;
292 sreq[creq]=Request::isend((char*)sendbuf +
293 segindex * segment_increment,
294 count_by_segment, datatype,
298 creq = (creq + 1) % max_outstanding_reqs;
300 original_count -= count_by_segment;
303 /* Wait on the remaining request to complete */
304 Request::waitall( max_outstanding_reqs, sreq,
305 MPI_STATUSES_IGNORE );
311 ompi_coll_tuned_topo_destroy_tree(&tree);
314 error_hndl: /* error handler */
315 XBT_DEBUG("ERROR_HNDL: node %d file %s line %d error %d\n",
316 rank, __FILE__, line, ret );
317 smpi_free_tmp_buffer(inbuf_free[0]);
318 smpi_free_tmp_buffer(inbuf_free[1]);
319 smpi_free_tmp_buffer(accumbuf);
323 /* Attention: this version of the reduce operations does not
325 - non-commutative operations
326 - segment sizes which are not multiplies of the extent of the datatype
327 meaning that at least one datatype must fit in the segment !
331 int reduce__ompi_chain(const void *sendbuf, void *recvbuf, int count,
332 MPI_Datatype datatype,
337 uint32_t segsize=64*1024;
338 int segcount = count;
340 int fanout = comm->size()/2;
342 XBT_DEBUG("coll:tuned:reduce_intra_chain rank %d fo %d ss %5u", comm->rank(), fanout, segsize);
345 * Determine number of segments and number of elements
348 typelng = datatype->size();
350 COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
352 return smpi_coll_tuned_ompi_reduce_generic( sendbuf, recvbuf, count, datatype,
354 ompi_coll_tuned_topo_build_chain(fanout, comm, root),
359 int reduce__ompi_pipeline(const void *sendbuf, void *recvbuf,
360 int count, MPI_Datatype datatype,
366 int segcount = count;
368 // COLL_TUNED_UPDATE_PIPELINE( comm, tuned_module, root );
371 * Determine number of segments and number of elements
374 const double a2 = 0.0410 / 1024.0; /* [1/B] */
375 const double b2 = 9.7128;
376 const double a4 = 0.0033 / 1024.0; /* [1/B] */
377 const double b4 = 1.6761;
378 typelng= datatype->size();
379 int communicator_size = comm->size();
380 size_t message_size = typelng * count;
382 if (communicator_size > (a2 * message_size + b2)) {
385 }else if (communicator_size > (a4 * message_size + b4)) {
393 XBT_DEBUG("coll:tuned:reduce_intra_pipeline rank %d ss %5u", comm->rank(), segsize);
395 COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
397 return smpi_coll_tuned_ompi_reduce_generic( sendbuf, recvbuf, count, datatype,
399 ompi_coll_tuned_topo_build_chain( 1, comm, root),
403 int reduce__ompi_binary(const void *sendbuf, void *recvbuf,
404 int count, MPI_Datatype datatype,
409 int segcount = count;
415 * Determine number of segments and number of elements
418 typelng=datatype->size();
423 XBT_DEBUG("coll:tuned:reduce_intra_binary rank %d ss %5u", comm->rank(), segsize);
425 COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
427 return smpi_coll_tuned_ompi_reduce_generic( sendbuf, recvbuf, count, datatype,
429 ompi_coll_tuned_topo_build_tree(2, comm, root),
433 int reduce__ompi_binomial(const void *sendbuf, void *recvbuf,
434 int count, MPI_Datatype datatype,
440 int segcount = count;
443 const double a1 = 0.6016 / 1024.0; /* [1/B] */
444 const double b1 = 1.3496;
446 // COLL_TUNED_UPDATE_IN_ORDER_BMTREE( comm, tuned_module, root );
449 * Determine number of segments and number of elements
452 typelng= datatype->size();
453 int communicator_size = comm->size();
454 size_t message_size = typelng * count;
455 if (((communicator_size < 8) && (message_size < 20480)) ||
456 (message_size < 2048) || (count <= 1)) {
459 } else if (communicator_size > (a1 * message_size + b1)) {
464 XBT_DEBUG("coll:tuned:reduce_intra_binomial rank %d ss %5u", comm->rank(), segsize);
465 COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
467 return smpi_coll_tuned_ompi_reduce_generic( sendbuf, recvbuf, count, datatype,
469 ompi_coll_tuned_topo_build_in_order_bmtree(comm, root),
474 * reduce_intra_in_order_binary
476 * Function: Logarithmic reduce operation for non-commutative operations.
477 * Accepts: same as MPI_Reduce()
478 * Returns: MPI_SUCCESS or error code
480 int reduce__ompi_in_order_binary(const void *sendbuf, void *recvbuf,
482 MPI_Datatype datatype,
488 int rank, size, io_root;
489 int segcount = count;
494 XBT_DEBUG("coll:tuned:reduce_intra_in_order_binary rank %d ss %5u", rank, segsize);
497 * Determine number of segments and number of elements
500 typelng=datatype->size();
501 COLL_TUNED_COMPUTED_SEGCOUNT( segsize, typelng, segcount );
503 /* An in-order binary tree must use root (size-1) to preserve the order of
504 operations. Thus, if root is not rank (size - 1), then we must handle
505 1. MPI_IN_PLACE option on real root, and
506 2. we must allocate temporary recvbuf on rank (size - 1).
507 Note that generic function must be careful not to switch order of
508 operations for non-commutative ops.
511 const void* use_this_sendbuf = sendbuf;
512 void* use_this_recvbuf = recvbuf;
513 unsigned char* tmp_sendbuf = nullptr;
514 unsigned char* tmp_recvbuf = nullptr;
515 if (io_root != root) {
518 ext=datatype->get_extent();
519 text=datatype->get_extent();
521 if ((root == rank) && (MPI_IN_PLACE == sendbuf)) {
522 tmp_sendbuf = smpi_get_tmp_sendbuffer(text + (count - 1) * ext);
523 if (NULL == tmp_sendbuf) {
524 return MPI_ERR_INTERN;
526 Datatype::copy(recvbuf, count, datatype, tmp_sendbuf, count, datatype);
527 use_this_sendbuf = tmp_sendbuf;
528 } else if (io_root == rank) {
529 tmp_recvbuf = smpi_get_tmp_recvbuffer(text + (count - 1) * ext);
530 if (NULL == tmp_recvbuf) {
531 return MPI_ERR_INTERN;
533 use_this_recvbuf = tmp_recvbuf;
537 /* Use generic reduce with in-order binary tree topology and io_root */
538 ret = smpi_coll_tuned_ompi_reduce_generic( use_this_sendbuf, use_this_recvbuf, count, datatype,
540 ompi_coll_tuned_topo_build_in_order_bintree(comm),
542 if (MPI_SUCCESS != ret) { return ret; }
545 if (io_root != root) {
547 /* Receive result from rank io_root to recvbuf */
548 Request::recv(recvbuf, count, datatype, io_root,
549 COLL_TAG_REDUCE, comm,
551 if (MPI_IN_PLACE == sendbuf) {
552 smpi_free_tmp_buffer(tmp_sendbuf);
555 } else if (io_root == rank) {
556 /* Send result from use_this_recvbuf to root */
557 Request::send(use_this_recvbuf, count, datatype, root,
560 smpi_free_tmp_buffer(tmp_recvbuf);
568 * Linear functions are copied from the BASIC coll module
569 * they do not segment the message and are simple implementations
570 * but for some small number of nodes and/or small data sizes they
571 * are just as fast as tuned/tree based segmenting operations
572 * and as such may be selected by the decision functions
573 * These are copied into this module due to the way we select modules
574 * in V1. i.e. in V2 we will handle this differently and so will not
575 * have to duplicate code.
576 * GEF Oct05 after asking Jeff.
579 /* copied function (with appropriate renaming) starts here */
584 * Function: - reduction using O(N) algorithm
585 * Accepts: - same as MPI_Reduce()
586 * Returns: - MPI_SUCCESS or error code
589 int reduce__ompi_basic_linear(const void *sbuf, void *rbuf, int count,
596 ptrdiff_t true_extent, lb, extent;
597 unsigned char* free_buffer = nullptr;
598 unsigned char* pml_buffer = nullptr;
599 unsigned char* inplace_temp = nullptr;
600 const unsigned char* inbuf;
607 XBT_DEBUG("coll:tuned:reduce_intra_basic_linear rank %d", rank);
609 /* If not root, send data to the root. */
612 Request::send(sbuf, count, dtype, root,
618 /* see discussion in ompi_coll_basic_reduce_lin_intra about
619 extent and true extent */
620 /* for reducing buffer allocation lengths.... */
622 dtype->extent(&lb, &extent);
623 true_extent = dtype->get_extent();
625 if (MPI_IN_PLACE == sbuf) {
627 inplace_temp = smpi_get_tmp_recvbuffer(true_extent + (count - 1) * extent);
628 if (nullptr == inplace_temp) {
631 rbuf = inplace_temp - lb;
635 free_buffer = smpi_get_tmp_recvbuffer(true_extent + (count - 1) * extent);
636 pml_buffer = free_buffer - lb;
639 /* Initialize the receive buffer. */
641 if (rank == (size - 1)) {
642 Datatype::copy((char*)sbuf, count, dtype,(char*)rbuf, count, dtype);
644 Request::recv(rbuf, count, dtype, size - 1,
645 COLL_TAG_REDUCE, comm,
649 /* Loop receiving and calling reduction function (C or Fortran). */
651 for (i = size - 2; i >= 0; --i) {
653 inbuf = static_cast<const unsigned char*>(sbuf);
655 Request::recv(pml_buffer, count, dtype, i,
656 COLL_TAG_REDUCE, comm,
661 /* Perform the reduction */
662 if(op!=MPI_OP_NULL) op->apply( inbuf, rbuf, &count, dtype);
665 if (nullptr != inplace_temp) {
666 Datatype::copy(inplace_temp, count, dtype, (char*)sbuf, count, dtype);
667 smpi_free_tmp_buffer(inplace_temp);
669 if (nullptr != free_buffer) {
670 smpi_free_tmp_buffer(free_buffer);
677 /* copied function (with appropriate renaming) ends here */