1 /* Copyright (c) 2013-2014. 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
21 /* -*- Mode: C; c-basic-offset:4 ; -*- */
22 /* Copyright (c) 2001-2014, The Ohio State University. All rights
25 * This file is part of the MVAPICH2 software package developed by the
26 * team members of The Ohio State University's Network-Based Computing
27 * Laboratory (NBCL), headed by Professor Dhabaleswar K. (DK) Panda.
29 * For detailed copyright and licensing information, please refer to the
30 * copyright file COPYRIGHT in the top level MVAPICH2 directory.
34 * (C) 2001 by Argonne National Laboratory.
35 * See COPYRIGHT in top-level directory.
37 #include "colls_private.h"
38 #include "src/smpi/smpi_group.hpp"
41 extern int (*MV2_Bcast_function) (void *buffer, int count, MPI_Datatype datatype,
42 int root, MPI_Comm comm_ptr);
44 extern int (*MV2_Bcast_intra_node_function) (void *buffer, int count, MPI_Datatype datatype,
45 int root, MPI_Comm comm_ptr);
47 extern int zcpy_knomial_factor;
48 extern int mv2_pipelined_zcpy_knomial_factor;
49 extern int bcast_segment_size;
50 extern int mv2_inter_node_knomial_factor;
51 extern int mv2_intra_node_knomial_factor;
52 extern int mv2_bcast_two_level_system_size;
53 #define INTRA_NODE_ROOT 0
55 #define MPIR_Pipelined_Bcast_Zcpy_MV2 smpi_coll_tuned_bcast_mpich
56 #define MPIR_Pipelined_Bcast_MV2 smpi_coll_tuned_bcast_mpich
57 #define MPIR_Bcast_binomial_MV2 smpi_coll_tuned_bcast_binomial_tree
58 #define MPIR_Bcast_scatter_ring_allgather_shm_MV2 smpi_coll_tuned_bcast_scatter_LR_allgather
59 #define MPIR_Bcast_scatter_doubling_allgather_MV2 smpi_coll_tuned_bcast_scatter_rdb_allgather
60 #define MPIR_Bcast_scatter_ring_allgather_MV2 smpi_coll_tuned_bcast_scatter_LR_allgather
61 #define MPIR_Shmem_Bcast_MV2 smpi_coll_tuned_bcast_mpich
62 #define MPIR_Bcast_tune_inter_node_helper_MV2 smpi_coll_tuned_bcast_mvapich2_inter_node
63 #define MPIR_Bcast_inter_node_helper_MV2 smpi_coll_tuned_bcast_mvapich2_inter_node
64 #define MPIR_Knomial_Bcast_intra_node_MV2 smpi_coll_tuned_bcast_mvapich2_knomial_intra_node
65 #define MPIR_Bcast_intra_MV2 smpi_coll_tuned_bcast_mvapich2_intra_node
67 extern int zcpy_knomial_factor;
68 extern int mv2_pipelined_zcpy_knomial_factor;
69 extern int bcast_segment_size;
70 extern int mv2_inter_node_knomial_factor;
71 extern int mv2_intra_node_knomial_factor;
72 #define mv2_bcast_two_level_system_size 64
73 #define mv2_bcast_short_msg 16384
74 #define mv2_bcast_large_msg 512*1024
75 #define mv2_knomial_intra_node_threshold 131072
76 #define mv2_scatter_rd_inter_leader_bcast 1
77 int smpi_coll_tuned_bcast_mvapich2_inter_node(void *buffer,
79 MPI_Datatype datatype,
84 int mpi_errno = MPI_SUCCESS;
85 MPI_Comm shmem_comm, leader_comm;
86 int local_rank, local_size, global_rank = -1;
87 int leader_root, leader_of_root;
91 //comm_size = comm->size();
94 if (MV2_Bcast_function==NULL){
95 MV2_Bcast_function=smpi_coll_tuned_bcast_mpich;
98 if (MV2_Bcast_intra_node_function==NULL){
99 MV2_Bcast_intra_node_function= smpi_coll_tuned_bcast_mpich;
102 if(comm->get_leaders_comm()==MPI_COMM_NULL){
106 shmem_comm = comm->get_intra_comm();
107 local_rank = shmem_comm->rank();
108 local_size = shmem_comm->size();
110 leader_comm = comm->get_leaders_comm();
112 if ((local_rank == 0) && (local_size > 1)) {
113 global_rank = leader_comm->rank();
116 int* leaders_map = comm->get_leaders_map();
117 leader_of_root = comm->group()->rank(leaders_map[root]);
118 leader_root = leader_comm->group()->rank(leaders_map[root]);
121 if (local_size > 1) {
122 if ((local_rank == 0) && (root != rank) && (leader_root == global_rank)) {
123 smpi_mpi_recv(buffer, count, datatype, root,
124 COLL_TAG_BCAST, comm, MPI_STATUS_IGNORE);
126 if ((local_rank != 0) && (root == rank)) {
127 smpi_mpi_send(buffer, count, datatype,
128 leader_of_root, COLL_TAG_BCAST, comm);
131 #if defined(_MCST_SUPPORT_)
132 if (comm_ptr->ch.is_mcast_ok) {
133 mpi_errno = MPIR_Mcast_inter_node_MV2(buffer, count, datatype, root, comm_ptr,
135 if (mpi_errno == MPI_SUCCESS) {
143 if (local_rank == 0) {
144 leader_comm = comm->get_leaders_comm();
148 if (MV2_Bcast_function == &MPIR_Pipelined_Bcast_MV2) {
149 mpi_errno = MPIR_Pipelined_Bcast_MV2(buffer, count, datatype,
151 } else if (MV2_Bcast_function == &MPIR_Bcast_scatter_ring_allgather_shm_MV2) {
152 mpi_errno = MPIR_Bcast_scatter_ring_allgather_shm_MV2(buffer, count,
156 if (local_rank == 0) {
157 /* if (MV2_Bcast_function == &MPIR_Knomial_Bcast_inter_node_wrapper_MV2) {
158 mpi_errno = MPIR_Knomial_Bcast_inter_node_wrapper_MV2(buffer, count,
162 mpi_errno = MV2_Bcast_function(buffer, count, datatype,
163 leader_root, leader_comm);
172 int smpi_coll_tuned_bcast_mvapich2_knomial_intra_node(void *buffer,
174 MPI_Datatype datatype,
175 int root, MPI_Comm comm)
177 int local_size = 0, rank;
178 int mpi_errno = MPI_SUCCESS;
179 MPI_Request *reqarray = NULL;
180 MPI_Status *starray = NULL;
181 int src, dst, mask, relative_rank;
183 if (MV2_Bcast_function==NULL){
184 MV2_Bcast_function=smpi_coll_tuned_bcast_mpich;
187 if (MV2_Bcast_intra_node_function==NULL){
188 MV2_Bcast_intra_node_function= smpi_coll_tuned_bcast_mpich;
191 if(comm->get_leaders_comm()==MPI_COMM_NULL){
195 local_size = comm->size();
199 reqarray=(MPI_Request *)xbt_malloc(2 * mv2_intra_node_knomial_factor * sizeof (MPI_Request));
201 starray=(MPI_Status *)xbt_malloc(2 * mv2_intra_node_knomial_factor * sizeof (MPI_Status));
203 /* intra-node k-nomial bcast */
204 if (local_size > 1) {
205 relative_rank = (rank >= root) ? rank - root : rank - root + local_size;
208 while (mask < local_size) {
209 if (relative_rank % (mv2_intra_node_knomial_factor * mask)) {
210 src = relative_rank / (mv2_intra_node_knomial_factor * mask) *
211 (mv2_intra_node_knomial_factor * mask) + root;
212 if (src >= local_size) {
216 smpi_mpi_recv(buffer, count, datatype, src,
217 COLL_TAG_BCAST, comm,
221 mask *= mv2_intra_node_knomial_factor;
223 mask /= mv2_intra_node_knomial_factor;
227 for (k = 1; k < mv2_intra_node_knomial_factor; k++) {
228 if (relative_rank + mask * k < local_size) {
229 dst = rank + mask * k;
230 if (dst >= local_size) {
233 reqarray[reqs++]=smpi_mpi_isend(buffer, count, datatype, dst,
234 COLL_TAG_BCAST, comm);
237 smpi_mpi_waitall(reqs, reqarray, starray);
239 mask /= mv2_intra_node_knomial_factor;
248 int smpi_coll_tuned_bcast_mvapich2_intra_node(void *buffer,
250 MPI_Datatype datatype,
251 int root, MPI_Comm comm)
253 int mpi_errno = MPI_SUCCESS;
255 int two_level_bcast = 1;
257 int is_homogeneous, is_contig;
259 void *tmp_buf = NULL;
264 if (MV2_Bcast_function==NULL){
265 MV2_Bcast_function=smpi_coll_tuned_bcast_mpich;
268 if (MV2_Bcast_intra_node_function==NULL){
269 MV2_Bcast_intra_node_function= smpi_coll_tuned_bcast_mpich;
272 if(comm->get_leaders_comm()==MPI_COMM_NULL){
276 comm_size = comm->size();
277 // rank = comm->rank();
279 if (HANDLE_GET_KIND(datatype) == HANDLE_KIND_BUILTIN)*/
282 MPID_Datatype_get_ptr(datatype, dtp);
283 is_contig = dtp->is_contig;
287 #ifdef MPID_HAS_HETERO
288 if (comm_ptr->is_hetero)
292 /* MPI_Type_size() might not give the accurate size of the packed
293 * datatype for heterogeneous systems (because of padding, encoding,
294 * etc). On the other hand, MPI_Pack_size() can become very
295 * expensive, depending on the implementation, especially for
296 * heterogeneous systems. We want to use MPI_Type_size() wherever
297 * possible, and MPI_Pack_size() in other places.
299 //if (is_homogeneous) {
300 type_size=smpi_datatype_size(datatype);
303 /* MPIR_Pack_size_impl(1, datatype, &type_size);*/
305 nbytes = (size_t) (count) * (type_size);
306 if (comm_size <= mv2_bcast_two_level_system_size) {
307 if (nbytes > mv2_bcast_short_msg && nbytes < mv2_bcast_large_msg) {
314 if (two_level_bcast == 1
315 #if defined(_MCST_SUPPORT_)
316 || comm_ptr->ch.is_mcast_ok
320 if (!is_contig || !is_homogeneous) {
321 tmp_buf=(void *)smpi_get_tmp_sendbuffer(nbytes);
323 /* TODO: Pipeline the packing and communication */
325 /* if (rank == root) {*/
327 /* MPIR_Pack_impl(buffer, count, datatype, tmp_buf, nbytes, &position);*/
329 /* MPIU_ERR_POP(mpi_errno);*/
333 shmem_comm = comm->get_intra_comm();
334 if (!is_contig || !is_homogeneous) {
336 MPIR_Bcast_inter_node_helper_MV2(tmp_buf, nbytes, MPI_BYTE,
340 MPIR_Bcast_inter_node_helper_MV2(buffer, count, datatype, root,
344 /* We are now done with the inter-node phase */
345 if (nbytes <= mv2_knomial_intra_node_threshold) {
346 if (!is_contig || !is_homogeneous) {
347 mpi_errno = MPIR_Shmem_Bcast_MV2(tmp_buf, nbytes, MPI_BYTE,
350 mpi_errno = MPIR_Shmem_Bcast_MV2(buffer, count, datatype,
354 if (!is_contig || !is_homogeneous) {
356 MPIR_Knomial_Bcast_intra_node_MV2(tmp_buf, nbytes,
362 MPIR_Knomial_Bcast_intra_node_MV2(buffer, count,
370 if (nbytes <= mv2_bcast_short_msg) {
371 mpi_errno = MPIR_Bcast_binomial_MV2(buffer, count, datatype, root,
374 if (mv2_scatter_rd_inter_leader_bcast) {
375 mpi_errno = MPIR_Bcast_scatter_ring_allgather_MV2(buffer, count,
381 MPIR_Bcast_scatter_doubling_allgather_MV2(buffer, count,