1 #include "colls_private.h"
3 /* change number of core per smp-node
4 we assume that number of core per process will be the same for all implementations */
10 This fucntion performs all-reduce operation as follow.
11 1) binomial_tree reduce inside each SMP node
12 2) reduce-scatter -inter between root of each SMP node
13 3) allgather - inter between root of each SMP node
14 4) binomial_tree bcast inside each SMP node
16 int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
17 int count, MPI_Datatype dtype, MPI_Op op,
22 int tag = COLL_TAG_ALLREDUCE;
25 int num_core = simcall_host_get_core(SIMIX_host_self());
26 // do we use the default one or the number of cores in the platform ?
27 // if the number of cores is one, the platform may be simulated with 1 node = 1 core
28 if (num_core == 1) num_core = NUM_CORE;
30 #ifdef MPICH2_REDUCTION
31 MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
33 MPI_User_function *uop;
34 struct MPIR_OP *op_ptr;
35 op_ptr = MPIR_ToPointer(op);
39 comm_size = smpi_comm_size(comm);
40 rank = smpi_comm_rank(comm);
42 extent = smpi_datatype_get_extent(dtype);
43 tmp_buf = (void *) xbt_malloc(count * extent);
45 int intra_rank, inter_rank;
46 intra_rank = rank % num_core;
47 inter_rank = rank / num_core;
49 //printf("node %d intra_rank = %d, inter_rank = %d\n", rank, intra_rank, inter_rank);
51 int inter_comm_size = (comm_size + num_core - 1) / num_core;
54 //printf("intra com size = %d\n",num_core);
55 //printf("inter com size = %d\n",inter_comm_size);
59 smpi_mpi_sendrecv(send_buf, count, dtype, rank, tag,
60 recv_buf, count, dtype, rank, tag, comm, &status);
63 // SMP_binomial_reduce
65 while (mask < num_core) {
66 if ((mask & intra_rank) == 0) {
67 src = (inter_rank * num_core) + (intra_rank | mask);
68 // if (src < ((inter_rank + 1) * num_core)) {
69 if (src < comm_size) {
70 smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
71 smpi_op_apply(op, tmp_buf, recv_buf, &count, &dtype);
72 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
76 dst = (inter_rank * num_core) + (intra_rank & (~mask));
77 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
78 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
86 // INTER: reduce-scatter
87 if (intra_rank == 0) {
88 int send_offset, recv_offset;
89 int seg_count = count / inter_comm_size;
90 int to = ((inter_rank + 1) % inter_comm_size) * num_core;
92 ((inter_rank + inter_comm_size - 1) % inter_comm_size) * num_core;
95 //printf("node %d to %d from %d\n",rank,to,from);
97 for (i = 0; i < (inter_comm_size - 1); i++) {
100 ((inter_rank - 1 - i +
101 inter_comm_size) % inter_comm_size) * seg_count * extent;
103 ((inter_rank - 2 - i +
104 inter_comm_size) % inter_comm_size) * seg_count * extent;
106 smpi_mpi_sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
107 tag + i, tmp_buf, seg_count, dtype, from, tag + i, comm,
111 smpi_op_apply(op, tmp_buf, (char *) recv_buf + recv_offset, &seg_count,
116 for (i = 0; i < (inter_comm_size - 1); i++) {
120 inter_comm_size) % inter_comm_size) * seg_count * extent;
122 ((inter_rank - 1 - i +
123 inter_comm_size) % inter_comm_size) * seg_count * extent;
125 smpi_mpi_sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
126 tag + i, (char *) recv_buf + recv_offset, seg_count, dtype,
127 from, tag + i, comm, &status);
135 // INTER_binomial_reduce
137 // only root node for each SMP
138 if (intra_rank == 0) {
141 while (mask < inter_comm_size) {
142 if ((mask & inter_rank) == 0) {
143 src = (inter_rank | mask) * num_core;
144 if (src < comm_size) {
145 smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
146 (* uop) (tmp_buf, recv_buf, &count, &dtype);
147 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
151 dst = (inter_rank & (~mask)) * num_core;
152 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
153 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
162 // INTER_binomial_bcast
165 if (intra_rank == 0) {
167 while (mask < inter_comm_size) {
168 if (inter_rank & mask) {
169 src = (inter_rank - mask) * num_core;
170 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
171 smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
178 //printf("My rank = %d my mask = %d\n", rank,mask);
181 if (inter_rank < inter_comm_size) {
182 dst = (inter_rank + mask) * num_core;
183 if (dst < comm_size) {
184 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
185 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
194 // INTRA_binomial_bcast
196 int num_core_in_current_smp = num_core;
197 if (inter_rank == (inter_comm_size - 1)) {
198 num_core_in_current_smp = comm_size - (inter_rank * num_core);
200 // printf("Node %d num_core = %d\n",rank, num_core_in_current_smp);
202 while (mask < num_core_in_current_smp) {
203 if (intra_rank & mask) {
204 src = (inter_rank * num_core) + (intra_rank - mask);
205 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
206 smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
213 //printf("My rank = %d my mask = %d\n", rank,mask);
216 dst = (inter_rank * num_core) + (intra_rank + mask);
217 if (dst < comm_size) {
218 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
219 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);