- int retval = MPI_SUCCESS;
-
- int index, rank;
- smpi_mpi_request_t request;
- int colorkey[2];
- smpi_mpi_status_t status;
-
- smpi_bench_end();
-
- // FIXME: need to test parameters
-
- index = smpi_host_index();
- rank = comm->index_to_rank_map[index];
-
- if (0 == rank) {
- int *colors = xbt_new(int, comm->size);
- int *keys = xbt_new(int, comm->size);
- int i, j, k;
- smpi_mpi_communicator_t tempcomm = NULL;
- int colortmp;
- int keycount;
- int *keystmp = xbt_new(int, comm->size);
- int *rankstmp = xbt_new(int, comm->size);
- int tmpval;
- int indextmp;
-
- colors[0] = color;
- keys[0] = key;
-
- // FIXME: not efficient
- for (i = 1; i < comm->size; i++) {
- retval = smpi_create_request(colorkey, 2, MPI_INT, MPI_ANY_SOURCE, rank, MPI_ANY_TAG, comm, &request);
- smpi_mpi_irecv(request);
- smpi_mpi_wait(request, &status);
- xbt_mallocator_release(smpi_global->request_mallocator, request);
- colors[i] = colorkey[0];
- keys[i] = colorkey[1];
- }
-
- for (i = 0; i < comm->size; i++) {
- if (-1 == colors[i]) {
- continue;
- }
- colortmp = colors[i];
- keycount = 0;
- for (j = i; j < comm->size; j++) {
- if(colortmp == colors[j]) {
- colors[j] = -1;
- keystmp[keycount] = keys[j];
- rankstmp[keycount] = j;
- keycount++;
- }
- }
- if (0 < keycount) {
- // FIXME: yes, mock me, bubble sort...
- for (j = 0; j < keycount; j++) {
- for (k = keycount - 1; k > j; k--) {
- if (keystmp[k] < keystmp[k - 1]) {
- tmpval = keystmp[k];
- keystmp[k] = keystmp[k - 1];
- keystmp[k - 1] = tmpval;
-
- tmpval = rankstmp[k];
- rankstmp[k] = rankstmp[k - 1];
- rankstmp[k - 1] = tmpval;
- }
- }
- }
- tempcomm = xbt_new(s_smpi_mpi_communicator_t, 1);
- tempcomm->barrier_count = 0;
- tempcomm->barrier_mutex = SIMIX_mutex_init();
- tempcomm->barrier_cond = SIMIX_cond_init();
- tempcomm->rank_to_index_map = xbt_new(int, keycount);
- tempcomm->index_to_rank_map = xbt_new(int, smpi_global->host_count);
- for (j = 0; j < smpi_global->host_count; j++) {
- tempcomm->index_to_rank_map[j] = -1;
- }
- for (j = 0; j < keycount; j++) {
- indextmp = comm->rank_to_index_map[rankstmp[j]];
- tempcomm->rank_to_index_map[j] = indextmp;
- tempcomm->index_to_rank_map[indextmp] = j;
- }
- for (j = 0; j < keycount; j++) {
- retval = smpi_create_request(&j, 1, MPI_INT, 0, rankstmp[j], 0, comm, &request);
- request->data = tempcomm;
- smpi_mpi_isend(request);
- smpi_mpi_wait(request, &status);
- xbt_mallocator_release(smpi_global->request_mallocator, request);
- }
- }
- }
- } else {
- colorkey[0] = color;
- colorkey[1] = key;
- retval = smpi_create_request(colorkey, 2, MPI_INT, rank, 0, 0, comm, &request);
- smpi_mpi_isend(request);
- smpi_mpi_wait(request, &status);
- xbt_mallocator_release(smpi_global->request_mallocator, request);
- retval = smpi_create_request(colorkey, 1, MPI_INT, 0, rank, 0, comm, &request);
- smpi_mpi_irecv(request);
- smpi_mpi_wait(request, &status);
- *comm_out = request->data;
- }
-
- smpi_bench_begin();
-
- return retval;
+ int retval = MPI_SUCCESS;
+
+ int index, rank;
+ smpi_mpi_request_t request;
+ int colorkey[2];
+ smpi_mpi_status_t status;
+
+ smpi_bench_end();
+
+ // FIXME: need to test parameters
+
+ index = smpi_host_index();
+ rank = comm->index_to_rank_map[index];
+
+ // default output
+ comm_out = NULL;
+
+ // root node does most of the real work
+ if (0 == rank) {
+ int colormap[comm->size];
+ int keymap[comm->size];
+ int rankkeymap[comm->size * 2];
+ int i, j;
+ smpi_mpi_communicator_t tempcomm = NULL;
+ int count;
+ int indextmp;
+
+ colormap[0] = color;
+ keymap[0] = key;
+
+ // FIXME: use scatter/gather or similar instead of individual comms
+ for (i = 1; i < comm->size; i++) {
+ retval = smpi_create_request(colorkey, 2, MPI_INT, MPI_ANY_SOURCE,
+ rank, MPI_ANY_TAG, comm, &request);
+ smpi_mpi_irecv(request);
+ smpi_mpi_wait(request, &status);
+ colormap[status.MPI_SOURCE] = colorkey[0];
+ keymap[status.MPI_SOURCE] = colorkey[1];
+ xbt_mallocator_release(smpi_global->request_mallocator, request);
+ }
+
+ for (i = 0; i < comm->size; i++) {
+ if (MPI_UNDEFINED == colormap[i]) {
+ continue;
+ }
+ // make a list of nodes with current color and sort by keys
+ count = 0;
+ for (j = i; j < comm->size; j++) {
+ if (colormap[i] == colormap[j]) {
+ colormap[j] = MPI_UNDEFINED;
+ rankkeymap[count * 2] = j;
+ rankkeymap[count * 2 + 1] = keymap[j];
+ count++;
+ }
+ }
+ qsort(rankkeymap, count, sizeof(int) * 2, &smpi_compare_rankkeys);
+
+ // new communicator
+ tempcomm = xbt_new(s_smpi_mpi_communicator_t, 1);
+ tempcomm->barrier_count = 0;
+ tempcomm->size = count;
+ tempcomm->barrier_mutex = SIMIX_mutex_init();
+ tempcomm->barrier_cond = SIMIX_cond_init();
+ tempcomm->rank_to_index_map = xbt_new(int, count);
+ tempcomm->index_to_rank_map = xbt_new(int, smpi_global->host_count);
+ for (j = 0; j < smpi_global->host_count; j++) {
+ tempcomm->index_to_rank_map[j] = -1;
+ }
+ for (j = 0; j < count; j++) {
+ indextmp = comm->rank_to_index_map[rankkeymap[j * 2]];
+ tempcomm->rank_to_index_map[j] = indextmp;
+ tempcomm->index_to_rank_map[indextmp] = j;
+ }
+ for (j = 0; j < count; j++) {
+ if (rankkeymap[j * 2]) {
+ retval = smpi_create_request(&j, 1, MPI_INT, 0,
+ rankkeymap[j * 2], 0, comm, &request);
+ request->data = tempcomm;
+ smpi_mpi_isend(request);
+ smpi_mpi_wait(request, &status);
+ xbt_mallocator_release(smpi_global->request_mallocator, request);
+ } else {
+ *comm_out = tempcomm;
+ }
+ }
+ }
+ } else {
+ colorkey[0] = color;
+ colorkey[1] = key;
+ retval = smpi_create_request(colorkey, 2, MPI_INT, rank, 0, 0, comm,
+ &request);
+ smpi_mpi_isend(request);
+ smpi_mpi_wait(request, &status);
+ xbt_mallocator_release(smpi_global->request_mallocator, request);
+ if (MPI_UNDEFINED != color) {
+ retval = smpi_create_request(colorkey, 1, MPI_INT, 0, rank, 0, comm,
+ &request);
+ smpi_mpi_irecv(request);
+ smpi_mpi_wait(request, &status);
+ *comm_out = request->data;
+ }
+ }
+
+ smpi_bench_begin();
+
+ return retval;