X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/d9179c8d05a4cda5b653afb5a45f398597919fb7..2ef76cc455c57d235189fdc567516e40f9c726d1:/src/smpi/smpi_mpi.c diff --git a/src/smpi/smpi_mpi.c b/src/smpi/smpi_mpi.c index 97ce5dfa92..97b50bb8bd 100644 --- a/src/smpi/smpi_mpi.c +++ b/src/smpi/smpi_mpi.c @@ -1,287 +1,450 @@ #include "private.h" -int MPI_Init(int *argc, char ***argv) +XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi, smpi, + "Logging specific to SMPI (mpi)"); + +int SMPI_MPI_Init(int *argc, char ***argv) { - smpi_mpi_init(); - smpi_bench_begin(); - return MPI_SUCCESS; + smpi_process_init(argc,argv); + smpi_bench_begin(); + return MPI_SUCCESS; } -int MPI_Finalize() +int SMPI_MPI_Finalize() { - smpi_bench_end(); - smpi_mpi_finalize(); - return MPI_SUCCESS; + smpi_bench_end(); + smpi_process_finalize(); + return MPI_SUCCESS; } // right now this just exits the current node, should send abort signal to all -// hosts in the communicator; -int MPI_Abort(MPI_Comm comm, int errorcode) +// hosts in the communicator (TODO) +int SMPI_MPI_Abort(MPI_Comm comm, int errorcode) { - smpi_exit(errorcode); - return 0; + smpi_exit(errorcode); + return 0; } -int MPI_Comm_size(MPI_Comm comm, int *size) +int SMPI_MPI_Comm_size(MPI_Comm comm, int *size) { - int retval = MPI_SUCCESS; + int retval = MPI_SUCCESS; - smpi_bench_end(); + smpi_bench_end(); - if (NULL == comm) { - retval = MPI_ERR_COMM; - } else if (NULL == size) { - retval = MPI_ERR_ARG; - } else { - *size = comm->size; - } + if (NULL == comm) { + retval = MPI_ERR_COMM; + } else if (NULL == size) { + retval = MPI_ERR_ARG; + } else { + *size = comm->size; + } - smpi_bench_begin(); + smpi_bench_begin(); - return retval; + return retval; } -int MPI_Comm_rank(MPI_Comm comm, int *rank) +int SMPI_MPI_Comm_rank(MPI_Comm comm, int *rank) { - int retval = MPI_SUCCESS; + int retval = MPI_SUCCESS; - smpi_bench_end(); + smpi_bench_end(); - if (NULL == comm) { - retval = MPI_ERR_COMM; - } else if (NULL == rank) { - retval = MPI_ERR_ARG; - } else { - *rank = smpi_mpi_comm_rank(comm); - } + if (NULL == comm) { + retval = MPI_ERR_COMM; + } else if (NULL == rank) { + retval = MPI_ERR_ARG; + } else { + *rank = smpi_mpi_comm_rank(comm); + } - smpi_bench_begin(); + smpi_bench_begin(); - return retval; + return retval; } -int MPI_Type_size(MPI_Datatype datatype, size_t *size) +int SMPI_MPI_Type_size(MPI_Datatype datatype, size_t * size) { - int retval = MPI_SUCCESS; + int retval = MPI_SUCCESS; - smpi_bench_end(); + smpi_bench_end(); - if (NULL == datatype) { - retval = MPI_ERR_TYPE; - } else if (NULL == size) { - retval = MPI_ERR_ARG; - } else { - *size = datatype->size; - } + if (NULL == datatype) { + retval = MPI_ERR_TYPE; + } else if (NULL == size) { + retval = MPI_ERR_ARG; + } else { + *size = datatype->size; + } - smpi_bench_begin(); + smpi_bench_begin(); - return retval; + return retval; } -int MPI_Barrier(MPI_Comm comm) +int SMPI_MPI_Barrier(MPI_Comm comm) { - int retval = MPI_SUCCESS; + int retval = MPI_SUCCESS; - smpi_bench_end(); + smpi_bench_end(); - if (NULL == comm) { - retval = MPI_ERR_COMM; - } else { - retval = smpi_mpi_barrier(comm); - } + if (NULL == comm) { + retval = MPI_ERR_COMM; + } else { + retval = smpi_mpi_barrier(comm); + } - smpi_bench_begin(); + smpi_bench_begin(); - return retval; + return retval; } -int MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Request *request) +int SMPI_MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src, + int tag, MPI_Comm comm, MPI_Request * request) { - int retval = MPI_SUCCESS; + int retval = MPI_SUCCESS; - smpi_bench_end(); + smpi_bench_end(); - if (NULL == request) { - retval = MPI_ERR_ARG; - } else { - int dst = 0; - retval = smpi_create_request(buf, count, datatype, src, dst, tag, comm, request); - if (NULL != *request && MPI_SUCCESS == retval) { - retval = smpi_mpi_irecv(*request); - } - } + retval = smpi_create_request(buf, count, datatype, src, 0, tag, comm, + request); + if (NULL != *request && MPI_SUCCESS == retval) { + retval = smpi_mpi_irecv(*request); + } - smpi_bench_begin(); + smpi_bench_begin(); - return retval; + return retval; } -int MPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status *status) +int SMPI_MPI_Recv(void *buf, int count, MPI_Datatype datatype, int src, + int tag, MPI_Comm comm, MPI_Status * status) { - int retval = MPI_SUCCESS; - int dst = 0; - smpi_mpi_request_t request; + int retval = MPI_SUCCESS; + smpi_mpi_request_t request; - smpi_bench_end(); + smpi_bench_end(); - retval = smpi_create_request(buf, count, datatype, src, dst, tag, comm, &request); - if (NULL != request && MPI_SUCCESS == retval) { - retval = smpi_mpi_irecv(request); - if (MPI_SUCCESS == retval) { - retval = smpi_mpi_wait(request, status); - } - xbt_mallocator_release(smpi_global->request_mallocator, request); - } + retval = smpi_create_request(buf, count, datatype, src, 0, tag, comm, + &request); + if (NULL != request && MPI_SUCCESS == retval) { + retval = smpi_mpi_irecv(request); + if (MPI_SUCCESS == retval) { + retval = smpi_mpi_wait(request, status); + } + xbt_mallocator_release(smpi_global->request_mallocator, request); + } - smpi_bench_begin(); + smpi_bench_begin(); - return retval; + return retval; } -int MPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm, MPI_Request *request) +int SMPI_MPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst, + int tag, MPI_Comm comm, MPI_Request * request) { - int retval = MPI_SUCCESS; + int retval = MPI_SUCCESS; - smpi_bench_end(); + smpi_bench_end(); - if (NULL == request) { - retval = MPI_ERR_ARG; - } else { - int src = 0; - retval = smpi_create_request(buf, count, datatype, src, dst, tag, comm, request); - if (NULL != *request && MPI_SUCCESS == retval) { - retval = smpi_mpi_isend(*request); - } - } + retval = smpi_create_request(buf, count, datatype, 0, dst, tag, comm, + request); + if (NULL != *request && MPI_SUCCESS == retval) { + retval = smpi_mpi_isend(*request); + } - smpi_bench_begin(); + smpi_bench_begin(); - return retval; + return retval; } -int MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) +int SMPI_MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, + int tag, MPI_Comm comm) { - int retval = MPI_SUCCESS; - int src = 0; - smpi_mpi_request_t request; + int retval = MPI_SUCCESS; + smpi_mpi_request_t request; - smpi_bench_end(); + smpi_bench_end(); - retval = smpi_create_request(buf, count, datatype, src, dst, tag, comm, &request); - if (NULL != request && MPI_SUCCESS == retval) { - retval = smpi_mpi_isend(request); - if (MPI_SUCCESS == retval) { - smpi_mpi_wait(request, MPI_STATUS_IGNORE); - } - xbt_mallocator_release(smpi_global->request_mallocator, request); - } + retval = smpi_create_request(buf, count, datatype, 0, dst, tag, comm, + &request); + if (NULL != request && MPI_SUCCESS == retval) { + retval = smpi_mpi_isend(request); + if (MPI_SUCCESS == retval) { + smpi_mpi_wait(request, MPI_STATUS_IGNORE); + } + xbt_mallocator_release(smpi_global->request_mallocator, request); + } - smpi_bench_begin(); + smpi_bench_begin(); - return retval; + return retval; } -int MPI_Comm_split(MPI_Comm comm, int color, int key, MPI_Comm *comm_out) +/** + * MPI_Wait and friends + **/ +int SMPI_MPI_Wait(MPI_Request * request, MPI_Status * status) { - 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; - } - // FIXME: now send new communicator to happy troops... + return smpi_mpi_wait(*request, status); +} + +int SMPI_MPI_Waitall(int count, MPI_Request requests[], MPI_Status status[]) { + return smpi_mpi_waitall(count, requests,status); +} +int SMPI_MPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status status[]) { + return smpi_mpi_waitany(count, requests, index,status); +} +/** + * MPI_Bcast + **/ +int SMPI_MPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root, + MPI_Comm comm) +{ + + int retval = MPI_SUCCESS; + int rank; + smpi_mpi_request_t request; + + smpi_bench_end(); + + rank = smpi_mpi_comm_rank(comm); + + if (rank == root) { + retval = smpi_create_request(buf, count, datatype, root, + (root + 1) % comm->size, 0, comm, &request); + request->forward = comm->size - 1; + smpi_mpi_isend(request); + } else { + retval = smpi_create_request(buf, count, datatype, MPI_ANY_SOURCE, rank, + 0, comm, &request); + smpi_mpi_irecv(request); + } + + smpi_mpi_wait(request, MPI_STATUS_IGNORE); + xbt_mallocator_release(smpi_global->request_mallocator, request); + + smpi_bench_begin(); + + return retval; +} + +void print_buffer_int( void *buf, int len, const char *msg) ; +void print_buffer_int( void *buf, int len, const char *msg) { + int tmp, *v; + printf("**%s: ",msg); + for (tmp=0;tmpsize; + + printf("-->rank %d. Entering ....\n",rank); + print_buffer_int( sendbuf, count, "sendbuf"); + + if (rank != root) { // if i am not ROOT, simply send my buffer to root + retval = smpi_create_request(sendbuf, count, datatype, rank, root, 0, comm, &request); + smpi_mpi_isend(request); + smpi_mpi_wait(request, MPI_STATUS_IGNORE); + xbt_mallocator_release(smpi_global->request_mallocator, request); + + } else { + // i am the ROOT: wait for all buffers by creating one request by sender + tabrequest = xbt_malloc((size-1)*sizeof(smpi_mpi_request_t)); + + void *tmprecvbuf = xbt_malloc(count*datatype->size); // to store intermediate receptions + memcpy(recvbuf,sendbuf,count*datatype->size*sizeof(char)); // initiliaze recv buf with my own snd buf + + // i can not use: 'request->forward = size-1;' (which would progagate size-1 receive reqs) + // since we should op values as soon as one receiving request matches. + for (i=0; isize-1; i++) { + // reminder: for smpi_create_request() the src is always the process sending. + retval = smpi_create_request(tmprecvbuf, count, datatype, MPI_ANY_SOURCE, root, + 0, comm, &(tabrequest[i])); + if (NULL != tabrequest[i] && MPI_SUCCESS == retval) { + if (MPI_SUCCESS == retval) { + smpi_mpi_irecv(tabrequest[i]); + } } - } + } + // now, wait for completion of all irecv's. + for (i=0; isize-1; i++) { + int index = MPI_UNDEFINED; + smpi_mpi_waitany(comm->size-1, tabrequest, &index, MPI_STATUS_IGNORE); + + print_buffer_int( recvbuf, count, "rcvbuf"); + printf("MPI_Waitany() unblocked: root received (completes req[%d]): ",index); + print_buffer_int( tmprecvbuf, count, "tmprecvbuf"); + + // arg 2 is modified + op->func (tmprecvbuf,recvbuf,&count,&datatype); + print_buffer_int( recvbuf, count, "recvbuf after func"); + //fprintf(stderr,"[smpi] %s:%d : MPI_Reduce *Not yet implemented*.\n",__FILE__,__LINE__); + xbt_mallocator_release(smpi_global->request_mallocator, tabrequest[i]); + } + xbt_free(tabrequest); + } + + smpi_bench_begin(); + + return retval; +} - } else { +// used by comm_split to sort ranks based on key values +int smpi_compare_rankkeys(const void *a, const void *b); +int smpi_compare_rankkeys(const void *a, const void *b) +{ + int *x = (int *) a; + int *y = (int *) b; + + if (x[1] < y[1]) + return -1; - 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 (x[1] == y[1]) { + if (x[0] < y[0]) + return -1; + if (x[0] == y[0]) + return 0; + return 1; + } - smpi_bench_begin(); + return 1; +} - return retval; +int SMPI_MPI_Comm_split(MPI_Comm comm, int color, int key, + MPI_Comm * comm_out) +{ + 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_process_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->process_count); + for (j = 0; j < smpi_global->process_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; +} + +double SMPI_MPI_Wtime( void ) +{ + return ( SIMIX_get_clock() ); }