X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/2ef76cc455c57d235189fdc567516e40f9c726d1..f0aa7a97261661fc3bf5c5ab9cb81128a3742bff:/src/smpi/smpi_mpi.c diff --git a/src/smpi/smpi_mpi.c b/src/smpi/smpi_mpi.c index 97b50bb8bd..4ee1d8a864 100644 --- a/src/smpi/smpi_mpi.c +++ b/src/smpi/smpi_mpi.c @@ -1,11 +1,14 @@ + + #include "private.h" +#include "smpi_coll_private.h" XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi, smpi, "Logging specific to SMPI (mpi)"); int SMPI_MPI_Init(int *argc, char ***argv) { - smpi_process_init(argc,argv); + smpi_process_init(argc, argv); smpi_bench_begin(); return MPI_SUCCESS; } @@ -85,13 +88,20 @@ int SMPI_MPI_Type_size(MPI_Datatype datatype, size_t * size) int SMPI_MPI_Barrier(MPI_Comm comm) { int retval = MPI_SUCCESS; + int arity=4; smpi_bench_end(); if (NULL == comm) { retval = MPI_ERR_COMM; } else { - retval = smpi_mpi_barrier(comm); + + /* + * original implemantation: + * retval = smpi_mpi_barrier(comm); + * this one is unrealistic: it just cond_waits, means no time. + */ + retval = nary_tree_barrier( comm, arity ); } smpi_bench_begin(); @@ -181,6 +191,52 @@ int SMPI_MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, return retval; } +/** + * MPI_Sendrecv + **/ +int SMPI_MPI_Sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, int dest, int sendtag, + void *recvbuf, int recvcount, MPI_Datatype recvtype, int source, int recvtag, + MPI_Comm comm, MPI_Status *status) +{ +int rank; +int retval = MPI_SUCCESS; +smpi_mpi_request_t srequest; +smpi_mpi_request_t rrequest; + + rank = smpi_mpi_comm_rank(comm); + + /* send */ + /* -------------*/ + retval = smpi_create_request(sendbuf, sendcount, sendtype, + rank,dest,sendtag, + comm, &srequest); + printf("[%d] isend request src=%d -> dst=%d (retval=%d)\n",rank,rank,dest,retval); + smpi_mpi_isend(srequest); + + + //retval = MPI_Isend( sendbuf, sendcount, sendtype, dest, sendtag, MPI_COMM_WORLD, &srequest); + + + /* recv */ + retval = smpi_create_request(recvbuf, recvcount, recvtype, + source, rank,recvtag, + comm, &rrequest); + printf("[%d] irecv request src=%d -> dst=%d (retval=%d)\n",rank,source,rank,retval); + smpi_mpi_irecv(rrequest); + + //retval = MPI_Irecv( recvbuf, recvcount, recvtype, source, recvtag, MPI_COMM_WORLD, &rrequest); + + + smpi_mpi_wait(srequest, MPI_STATUS_IGNORE); + printf("[%d] isend request src=%d dst=%d tag=%d COMPLETED (retval=%d) \n",rank,rank,dest,sendtag,retval); + + smpi_mpi_wait(rrequest, MPI_STATUS_IGNORE); + printf("[%d] irecv request src=%d -> dst=%d tag=%d COMPLETED (retval=%d)\n",rank,source,rank,recvtag,retval); + + return(retval); +} + + /** * MPI_Wait and friends **/ @@ -189,122 +245,340 @@ int SMPI_MPI_Wait(MPI_Request * request, MPI_Status * status) 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_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); + +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, + +/** + * flat bcast + **/ +int flat_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm); +int flat_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root, + MPI_Comm comm) +{ + int rank; + int retval = MPI_SUCCESS; + smpi_mpi_request_t request; + + 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); + + return(retval); + +} +/** + * Bcast internal level + **/ +int smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm) { + int retval = MPI_SUCCESS; + //retval = flat_tree_bcast(buf, count, datatype, root, comm); + retval = nary_tree_bcast(buf, count, datatype, root, comm, 2 ); + return retval; +} +/** + * Bcast user entry point + **/ +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(); + smpi_mpi_bcast(buf,count,datatype,root,comm); + smpi_bench_begin(); - rank = smpi_mpi_comm_rank(comm); + return retval; +} - 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); + + +#ifdef DEBUG_REDUCE +/** + * debugging helper function + **/ +static void print_buffer_int(void *buf, int len, char *msg, int rank) +{ + int tmp, *v; + printf("**[%d] %s: ", rank, msg); + for (tmp = 0; tmp < len; tmp++) { + v = buf; + printf("[%d]", v[tmp]); } + printf("\n"); + free(msg); +} +static void print_buffer_double(void *buf, int len, char *msg, int rank) +{ + int tmp; + double *v; + printf("**[%d] %s: ", rank, msg); + for (tmp = 0; tmp < len; tmp++) { + v = buf; + printf("[%lf]", v[tmp]); + } + printf("\n"); + free(msg); +} - smpi_mpi_wait(request, MPI_STATUS_IGNORE); - xbt_mallocator_release(smpi_global->request_mallocator, request); - smpi_bench_begin(); +#endif +/** + * MPI_Reduce + **/ +int smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, + MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm) +{ + int retval = MPI_SUCCESS; + int rank; + int size; + int i; + int tag = 0; + smpi_mpi_request_t *requests; + smpi_mpi_request_t request; + + smpi_bench_end(); + + rank = smpi_mpi_comm_rank(comm); + size = comm->size; + + if (rank != root) { // if i am not ROOT, simply send my buffer to root + +#ifdef DEBUG_REDUCE + print_buffer_int(sendbuf, count, xbt_strdup("sndbuf"), rank); +#endif + retval = smpi_create_request(sendbuf, count, datatype, rank, root, tag, comm, + &request); + smpi_mpi_isend(request); + smpi_mpi_wait(request, MPI_STATUS_IGNORE); + xbt_mallocator_release(smpi_global->request_mallocator, request); - return retval; + } else { + // i am the ROOT: wait for all buffers by creating one request by sender + int src; + requests = xbt_malloc((size-1) * sizeof(smpi_mpi_request_t)); + + void **tmpbufs = xbt_malloc((size-1) * sizeof(void *)); + for (i = 0; i < size-1; i++) { + // we need 1 buffer per request to store intermediate receptions + tmpbufs[i] = xbt_malloc(count * datatype->size); + } + // root: initiliaze recv buf with my own snd buf + memcpy(recvbuf, sendbuf, count * datatype->size * sizeof(char)); + + // 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; i < size-1; i++) { + // reminder: for smpi_create_request() the src is always the process sending. + src = i < root ? i : i + 1; + retval = smpi_create_request(tmpbufs[i], count, datatype, + src, root, tag, comm, &(requests[i])); + if (NULL != requests[i] && MPI_SUCCESS == retval) { + if (MPI_SUCCESS == retval) { + smpi_mpi_irecv(requests[i]); + } + } + } + // now, wait for completion of all irecv's. + for (i = 0; i < size-1; i++) { + int index = MPI_UNDEFINED; + smpi_mpi_waitany( size-1, requests, &index, MPI_STATUS_IGNORE); +#ifdef DEBUG_REDUCE + printf ("MPI_Waitany() unblocked: root received (completes req[index=%d])\n",index); + print_buffer_int(tmpbufs[index], count, bprintf("tmpbufs[index=%d] (value received)", index), + rank); +#endif + + // arg 2 is modified + op->func(tmpbufs[index], recvbuf, &count, &datatype); +#ifdef DEBUG_REDUCE + print_buffer_int(recvbuf, count, xbt_strdup("rcvbuf"), rank); +#endif + xbt_free(tmpbufs[index]); + /* FIXME: with the following line, it generates an + * [xbt_ex/CRITICAL] Conditional list not empty 162518800. + */ + // xbt_mallocator_release(smpi_global->request_mallocator, requests[index]); + } + xbt_free(requests); + xbt_free(tmpbufs); + } + 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]); + requests = xbt_malloc((comm->size-1) * sizeof(smpi_mpi_request_t)); + if (rank == root) { + // i am the root: distribute my sendbuf + //print_buffer_int(sendbuf, comm->size, xbt_strdup("rcvbuf"), rank); + cptr = sendbuf; + for (i=0; i < comm->size; i++) { + if ( i!=root ) { // send to processes ... + + retval = smpi_create_request((void *)cptr, sendcount, + datatype, root, i, tag, comm, &(requests[cnt])); + if (NULL != requests[cnt] && MPI_SUCCESS == retval) { + if (MPI_SUCCESS == retval) { + smpi_mpi_isend(requests[cnt]); + } } + cnt++; + } + else { // ... except if it's me. + memcpy(recvbuf, (void *)cptr, recvcount*recvtype->size*sizeof(char)); } + cptr += sendcount*datatype->size; + } + for(i=0; irequest_mallocator, requests[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]); + } + else { // i am a non-root process: wait data from the root + retval = smpi_create_request(recvbuf,recvcount, + recvtype, root, rank, tag, comm, &request); + if (NULL != request && MPI_SUCCESS == retval) { + if (MPI_SUCCESS == retval) { + smpi_mpi_irecv(request); + } } - xbt_free(tabrequest); + smpi_mpi_wait(request, &status); + xbt_mallocator_release(smpi_global->request_mallocator, request); + } + xbt_free(requests); + + smpi_bench_begin(); + + return retval; +} + + +/** + * MPI_Alltoall user entry point + * + * Uses the logic of OpenMPI (upto 1.2.7 or greater) for the optimizations + * ompi/mca/coll/tuned/coll_tuned_module.c + **/ +int SMPI_MPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype datatype, + void *recvbuf, int recvcount, MPI_Datatype recvtype, + MPI_Comm comm) +{ + int retval = MPI_SUCCESS; + int block_dsize; + int rank; + + smpi_bench_end(); + + rank = smpi_mpi_comm_rank(comm); + block_dsize = datatype->size * sendcount; + + if ((block_dsize < 200) && (comm->size > 12)) { + retval = smpi_coll_tuned_alltoall_bruck(sendbuf, sendcount, datatype, + recvbuf, recvcount, recvtype, comm); + + } else if (block_dsize < 3000) { +/* use this one !! retval = smpi_coll_tuned_alltoall_basic_linear(sendbuf, sendcount, datatype, + recvbuf, recvcount, recvtype, comm); + */ + retval = smpi_coll_tuned_alltoall_pairwise(sendbuf, sendcount, datatype, + recvbuf, recvcount, recvtype, comm); + } else { + + retval = smpi_coll_tuned_alltoall_pairwise(sendbuf, sendcount, datatype, + recvbuf, recvcount, recvtype, comm); } smpi_bench_begin(); @@ -312,12 +586,15 @@ int SMPI_MPI_Reduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datat return retval; } + + + // 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; + int *x = (int *) a; + int *y = (int *) b; if (x[1] < y[1]) return -1; @@ -444,7 +721,9 @@ int SMPI_MPI_Comm_split(MPI_Comm comm, int color, int key, return retval; } -double SMPI_MPI_Wtime( void ) +double SMPI_MPI_Wtime(void) { - return ( SIMIX_get_clock() ); + return (SIMIX_get_clock()); } + +