X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/a468c44a3ff64c8252a45ba0c4d8fea2c147491b..46a70e098d1d50876b9cf0576cac2357725cf54f:/src/smpi/smpi_base.c diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c index d7d32fcdd3..84c53dc403 100644 --- a/src/smpi/smpi_base.c +++ b/src/smpi/smpi_base.c @@ -12,7 +12,7 @@ #include "simix/smx_private.h" #include "surf/surf.h" #include "simgrid/sg_config.h" - +#include "colls/colls.h" XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_base, smpi, "Logging specific to SMPI (base)"); @@ -197,7 +197,7 @@ static MPI_Request build_request(void *buf, int count, if(datatype->has_subtype == 1){ // This part handles the problem of non-contiguous memory old_buf = buf; - buf = xbt_malloc(count*smpi_datatype_size(datatype)); + buf = count==0 ? NULL : xbt_malloc(count*smpi_datatype_size(datatype)); if (flags & SEND) { subtype->serialize(old_buf, buf, count, datatype->substruct); } @@ -284,7 +284,7 @@ MPI_Request smpi_mpi_send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = - build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, + build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SEND); request->refcount++; return request; @@ -294,7 +294,7 @@ MPI_Request smpi_mpi_ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = - build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, + build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, comm, PERSISTENT | SSEND | SEND); request->refcount++; return request; @@ -304,7 +304,7 @@ MPI_Request smpi_mpi_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm) { MPI_Request request = - build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag, + build_request(buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, comm, PERSISTENT | RECV); request->refcount++; return request; @@ -337,7 +337,7 @@ void smpi_mpi_start(MPI_Request request) } else { - int receiver = smpi_group_index(smpi_comm_group(request->comm), request->dst); + int receiver = request->dst;//smpi_group_index(smpi_comm_group(request->comm), request->dst); #ifdef HAVE_TRACING int rank = smpi_process_index(); @@ -362,13 +362,14 @@ void smpi_mpi_start(MPI_Request request) request->refcount++; if(request->old_type->has_subtype == 0){ oldbuf = request->buf; - if (oldbuf){ + if (oldbuf && request->size!=0){ request->buf = xbt_malloc(request->size); memcpy(request->buf,oldbuf,request->size); } } XBT_DEBUG("Send request %p is detached; buf %p copied into %p",request,oldbuf,request->buf); } + // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later request->real_size=request->size; smpi_datatype_use(request->old_type); @@ -422,8 +423,12 @@ void smpi_mpi_request_free(MPI_Request * request) if((*request)->refcount<0) xbt_die("wrong refcount"); if((*request)->refcount==0){ + print_request("Destroying", (*request)); xbt_free(*request); *request = MPI_REQUEST_NULL; + }else{ + print_request("Decrementing", (*request)); + } }else{ xbt_die("freeing an already free request"); @@ -434,7 +439,7 @@ MPI_Request smpi_isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = - build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, + build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SEND); return request; @@ -444,7 +449,7 @@ MPI_Request smpi_mpi_isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = - build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, + build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | ISEND | SEND); smpi_mpi_start(request); @@ -455,7 +460,7 @@ MPI_Request smpi_mpi_issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = - build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, + build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | ISEND | SSEND | SEND); smpi_mpi_start(request); return request; @@ -467,7 +472,7 @@ MPI_Request smpi_irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm) { MPI_Request request = - build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag, + build_request(buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, comm, NON_PERSISTENT | RECV); return request; } @@ -476,7 +481,7 @@ MPI_Request smpi_mpi_irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm) { MPI_Request request = - build_request(buf, count, datatype, src, smpi_comm_rank(comm), tag, + build_request(buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), src), smpi_process_index(), tag, comm, NON_PERSISTENT | RECV); smpi_mpi_start(request); @@ -497,9 +502,8 @@ void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { MPI_Request request = - build_request(buf, count, datatype, smpi_comm_rank(comm), dst, tag, + build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, comm, NON_PERSISTENT | SEND); - smpi_mpi_start(request); smpi_mpi_wait(&request, MPI_STATUS_IGNORE); @@ -508,7 +512,11 @@ void smpi_mpi_send(void *buf, int count, MPI_Datatype datatype, int dst, void smpi_mpi_ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm) { - MPI_Request request = smpi_mpi_issend(buf, count, datatype, dst, tag, comm); + MPI_Request request = + build_request(buf, count, datatype, smpi_process_index(), smpi_group_index(smpi_comm_group(comm), dst), tag, + comm, NON_PERSISTENT | SSEND | SEND); + + smpi_mpi_start(request); smpi_mpi_wait(&request, MPI_STATUS_IGNORE); } @@ -519,7 +527,12 @@ void smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, { MPI_Request requests[2]; MPI_Status stats[2]; - + int myid=smpi_process_index(); + if ((dst == myid) && (src == myid)) { + smpi_datatype_copy(sendbuf, sendcount, sendtype, + recvbuf, recvcount, recvtype); + return; + } requests[0] = smpi_isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm); requests[1] = @@ -545,7 +558,8 @@ static void finish_wait(MPI_Request * request, MPI_Status * status) if(!(req->detached && req->flags & SEND)){ if(status != MPI_STATUS_IGNORE) { - status->MPI_SOURCE = req->src == MPI_ANY_SOURCE ? req->real_src : req->src; + int src = req->src == MPI_ANY_SOURCE ? req->real_src : req->src; + status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(req->comm), src); status->MPI_TAG = req->tag == MPI_ANY_TAG ? req->real_tag : req->tag; status->MPI_ERROR = req->truncated ? MPI_ERR_TRUNCATE : MPI_SUCCESS; // this handles the case were size in receive differs from size in send @@ -678,16 +692,21 @@ void smpi_mpi_probe(int source, int tag, MPI_Comm comm, MPI_Status* status){ while(flag==0){ smpi_mpi_iprobe(source, tag, comm, &flag, status); XBT_DEBUG("Busy Waiting on probing : %d", flag); - if(!flag) { - simcall_process_sleep(0.0001); - } } } void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){ - MPI_Request request =build_request(NULL, 0, MPI_CHAR, source, smpi_comm_rank(comm), tag, + + MPI_Request request =build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag, comm, NON_PERSISTENT | RECV); + //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls + double sleeptime= sg_cfg_get_double("smpi/iprobe"); + //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it + static int nsleeps = 1; + + simcall_process_sleep(sleeptime); + // behave like a receive, but don't do it smx_rdv_t mailbox; @@ -708,13 +727,17 @@ void smpi_mpi_iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action); *flag = 1; if(status != MPI_STATUS_IGNORE) { - status->MPI_SOURCE = req->src; + status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src); status->MPI_TAG = req->tag; status->MPI_ERROR = MPI_SUCCESS; status->count = req->real_size; } + nsleeps=1;//reset the number of sleeps we will do next time + } + else { + *flag = 0; + nsleeps++; } - else *flag = 0; smpi_mpi_request_free(&request); return; @@ -923,6 +946,31 @@ void smpi_mpi_gather(void *sendbuf, int sendcount, MPI_Datatype sendtype, } } + +void smpi_mpi_reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, + MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) +{ + int i, size, count; + int *displs; + int rank = smpi_process_index(); + void *tmpbuf; + + /* arbitrarily choose root as rank 0 */ + size = smpi_comm_size(comm); + count = 0; + displs = xbt_new(int, size); + for (i = 0; i < size; i++) { + displs[i] = count; + count += recvcounts[i]; + } + tmpbuf=(void*)xbt_malloc(count*smpi_datatype_get_extent(datatype)); + mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm); + smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf, + recvcounts[rank], datatype, 0, comm); + xbt_free(displs); + xbt_free(tmpbuf); +} + void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype, void *recvbuf, int *recvcounts, int *displs, MPI_Datatype recvtype, int root, MPI_Comm comm) @@ -1128,17 +1176,29 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count, MPI_Request *requests; void **tmpbufs; + char* sendtmpbuf = (char*) sendbuf; + if( sendbuf == MPI_IN_PLACE ) { + sendtmpbuf = (char *)recvbuf; + } + rank = smpi_comm_rank(comm); size = smpi_comm_size(comm); + //non commutative case, use a working algo from openmpi + if(!smpi_op_is_commute(op)){ + smpi_coll_tuned_reduce_ompi_basic_linear(sendtmpbuf, recvbuf, count, + datatype, op, root, comm); + return; + } + if(rank != root) { // Send buffer to root - smpi_mpi_send(sendbuf, count, datatype, root, system_tag, comm); + smpi_mpi_send(sendtmpbuf, count, datatype, root, system_tag, comm); } else { // FIXME: check for errors smpi_datatype_extent(datatype, &lb, &dataext); // Local copy from root - if (sendbuf && recvbuf) - smpi_datatype_copy(sendbuf, count, datatype, recvbuf, count, datatype); + if (sendtmpbuf && recvbuf) + smpi_datatype_copy(sendtmpbuf, count, datatype, recvbuf, count, datatype); // Receive buffers from senders //TODO: make a MPI_barrier here ? requests = xbt_new(MPI_Request, size - 1);