1 /* Copyright (c) 2013-2022. 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. */
7 #include "../colls_private.hpp"
8 /* IMPLEMENTED BY PITCH PATARASUK
9 Non-topology-specific (however, number of cores/node need to be changed)
10 all-reduce operation designed for smp clusters
11 It uses 2-layer communication: binomial for intra-communication
12 and rdb for inter-communication*/
15 Use -DMPICH2 if this code does not compile.
16 MPICH1 code also work on MPICH2 on our cluster and the performance are similar.
17 This code assume commutative and associative reduce operator (MPI_SUM, MPI_MAX, etc).
20 //#include <star-reduction.c>
23 This function performs all-reduce operation as follow.
24 1) binomial_tree reduce inside each SMP node
25 2) Recursive doubling intra-communication between root of each SMP node
26 3) binomial_tree bcast inside each SMP node
30 int allreduce__smp_rdb(const void *send_buf, void *recv_buf, int count,
31 MPI_Datatype dtype, MPI_Op op,
35 int tag = COLL_TAG_ALLREDUCE;
38 if(comm->get_leaders_comm()==MPI_COMM_NULL){
42 if (comm->is_uniform()){
43 num_core = comm->get_intra_comm()->size();
46 #ifdef MPICH2_REDUCTION
47 MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
49 MPI_User_function *uop;
51 op_ptr = MPIR_ToPointer(op);
55 comm_size = comm->size();
58 extent = dtype->get_extent();
59 unsigned char* tmp_buf = smpi_get_tmp_sendbuffer(count * extent);
61 /* compute intra and inter ranking */
62 int intra_rank, inter_rank;
63 intra_rank = rank % num_core;
64 inter_rank = rank / num_core;
66 /* size of processes participate in intra communications =>
67 should be equal to number of machines */
68 int inter_comm_size = (comm_size + num_core - 1) / num_core;
70 /* copy input buffer to output buffer */
71 Request::sendrecv(send_buf, count, dtype, rank, tag,
72 recv_buf, count, dtype, rank, tag, comm, &status);
74 /* start binomial reduce intra communication inside each SMP node */
76 while (mask < num_core) {
77 if ((mask & intra_rank) == 0) {
78 src = (inter_rank * num_core) + (intra_rank | mask);
79 if (src < comm_size) {
80 Request::recv(tmp_buf, count, dtype, src, tag, comm, &status);
81 if(op!=MPI_OP_NULL) op->apply( tmp_buf, recv_buf, &count, dtype);
84 dst = (inter_rank * num_core) + (intra_rank & (~mask));
85 Request::send(recv_buf, count, dtype, dst, tag, comm);
89 } /* end binomial reduce intra-communication */
92 /* start rdb (recursive doubling) all-reduce inter-communication
93 between each SMP nodes : each node only have one process that can communicate
95 if (intra_rank == 0) {
97 /* find nearest power-of-two less than or equal to inter_comm_size */
98 int pof2, rem, newrank, newdst;
100 while (pof2 <= inter_comm_size)
103 rem = inter_comm_size - pof2;
105 /* In the non-power-of-two case, all even-numbered
106 processes of rank < 2*rem send their data to
107 (rank+1). These even-numbered processes no longer
108 participate in the algorithm until the very end.
110 if (inter_rank < 2 * rem) {
111 if (inter_rank % 2 == 0) {
112 dst = rank + num_core;
113 Request::send(recv_buf, count, dtype, dst, tag, comm);
116 src = rank - num_core;
117 Request::recv(tmp_buf, count, dtype, src, tag, comm, &status);
118 if(op!=MPI_OP_NULL) op->apply( tmp_buf, recv_buf, &count, dtype);
119 newrank = inter_rank / 2;
122 newrank = inter_rank - rem;
125 /* example inter-communication RDB rank change algorithm
126 0,4,8,12..36 <= true rank (assume 4 core per SMP)
127 0123 4567 89 <= inter_rank
128 1 3 4567 89 (1,3 got data from 0,2 : 0,2 will be idle until the end)
130 0 1 2345 67 => newrank
135 while (mask < pof2) {
136 newdst = newrank ^ mask;
137 /* find real rank of dest */
138 dst = (newdst < rem) ? newdst * 2 + 1 : newdst + rem;
141 /* exchange data in rdb manner */
142 Request::sendrecv(recv_buf, count, dtype, dst, tag, tmp_buf, count, dtype,
143 dst, tag, comm, &status);
144 if(op!=MPI_OP_NULL) op->apply( tmp_buf, recv_buf, &count, dtype);
150 left-over processes (all even ranks: < 2 * rem) get the result
152 if (inter_rank < 2 * rem) {
153 if (inter_rank % 2) {
154 Request::send(recv_buf, count, dtype, rank - num_core, tag, comm);
156 Request::recv(recv_buf, count, dtype, rank + num_core, tag, comm, &status);
161 /* start binomial broadcast intra-communication inside each SMP nodes */
162 int num_core_in_current_smp = num_core;
163 if (inter_rank == (inter_comm_size - 1)) {
164 num_core_in_current_smp = comm_size - (inter_rank * num_core);
167 while (mask < num_core_in_current_smp) {
168 if (intra_rank & mask) {
169 src = (inter_rank * num_core) + (intra_rank - mask);
170 Request::recv(recv_buf, count, dtype, src, tag, comm, &status);
178 dst = (inter_rank * num_core) + (intra_rank + mask);
179 if (dst < comm_size) {
180 Request::send(recv_buf, count, dtype, dst, tag, comm);
185 smpi_free_tmp_buffer(tmp_buf);