X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/d38fb65942fcf1a4bd05cb86c0b86b7bbbd3a47f..f18ab288219690177e48123c64ba809bd421f02a:/src/smpi/smpi_coll.c diff --git a/src/smpi/smpi_coll.c b/src/smpi/smpi_coll.c index 54f1e2db8c..d1984b0a26 100644 --- a/src/smpi/smpi_coll.c +++ b/src/smpi/smpi_coll.c @@ -1,12 +1,10 @@ -/* $Id$tag */ - /* smpi_coll.c -- various optimized routing for collectives */ -/* Copyright (c) 2009 Stephane Genaud. */ -/* All rights reserved. */ +/* Copyright (c) 2009, 2010. The SimGrid Team. + * All rights reserved. */ /* This program is free software; you can redistribute it and/or modify it - * * under the terms of the license (GNU LGPL) which comes with this package. */ + * under the terms of the license (GNU LGPL) which comes with this package. */ #include #include @@ -38,7 +36,7 @@ static proc_tree_t alloc_tree(int arity) { tree = xbt_new(struct s_proc_tree, 1); tree->PROCTREE_A = arity; - tree->isRoot = 0; + tree->isRoot = 0; tree->numChildren = 0; tree->child = xbt_new(int, arity); for(i = 0; i < arity; i++) { @@ -102,16 +100,17 @@ static void tree_bcast(void* buf, int count, MPI_Datatype datatype, int root, MP smpi_mpi_recv(buf, count, datatype, tree->parent, system_tag + rank, comm, MPI_STATUS_IGNORE); } requests = xbt_new(MPI_Request, tree->numChildren); - DEBUG2("<%d> creates %d requests (1 per child)\n", rank, tree->numChildren); + DEBUG2("<%d> creates %d requests (1 per child)", rank, tree->numChildren); /* iniates sends to ranks lower in the tree */ for(i = 0; i < tree->numChildren; i++) { if(tree->child[i] == -1) { requests[i] = MPI_REQUEST_NULL; } else { DEBUG3("<%d> send to <%d>, tag=%d", rank, tree->child[i], system_tag + tree->child[i]); - requests[i] = smpi_mpi_isend(buf, count, datatype, tree->child[i], system_tag + tree->child[i], comm); + requests[i] = smpi_isend_init(buf, count, datatype, tree->child[i], system_tag + tree->child[i], comm); } } + smpi_mpi_startall(tree->numChildren, requests); smpi_mpi_waitall(tree->numChildren, requests, MPI_STATUS_IGNORE); xbt_free(requests); } @@ -133,24 +132,25 @@ static void tree_antibcast(void* buf, int count, MPI_Datatype datatype, int root } //every one receives as many messages as it has children requests = xbt_new(MPI_Request, tree->numChildren); - DEBUG2("<%d> creates %d requests (1 per child)\n", rank, tree->numChildren); + DEBUG2("<%d> creates %d requests (1 per child)", rank, tree->numChildren); for(i = 0; i < tree->numChildren; i++) { if(tree->child[i] == -1) { requests[i] = MPI_REQUEST_NULL; } else { DEBUG3("<%d> recv from <%d>, tag=%d", rank, tree->child[i], system_tag + tree->child[i]); - requests[i] = smpi_mpi_irecv(buf, count, datatype, tree->child[i], system_tag + tree->child[i], comm); + requests[i] = smpi_irecv_init(buf, count, datatype, tree->child[i], system_tag + tree->child[i], comm); } } + smpi_mpi_startall(tree->numChildren, requests); smpi_mpi_waitall(tree->numChildren, requests, MPI_STATUS_IGNORE); xbt_free(requests); -} +} /** - * bcast with a binary, ternary, or whatever tree .. + * bcast with a binary, ternary, or whatever tree .. **/ void nary_tree_bcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm, int arity) { - proc_tree_t tree = alloc_tree(arity); + proc_tree_t tree = alloc_tree(arity); int rank, size; rank = smpi_comm_rank(comm); @@ -161,10 +161,10 @@ void nary_tree_bcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_ } /** - * barrier with a binary, ternary, or whatever tree .. + * barrier with a binary, ternary, or whatever tree .. **/ void nary_tree_barrier(MPI_Comm comm, int arity) { - proc_tree_t tree = alloc_tree( arity ); + proc_tree_t tree = alloc_tree( arity ); int rank, size; char dummy='$'; @@ -177,12 +177,52 @@ void nary_tree_barrier(MPI_Comm comm, int arity) { } /** - * Alltoall Bruck + * Alltoall Bruck * * Openmpi calls this routine when the message size sent to each rank < 2000 bytes and size < 12 **/ int smpi_coll_tuned_alltoall_bruck(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) { - DEBUG0("coll:tuned:alltoall_intra_bruck ** NOT IMPLEMENTED YET**"); + int system_tag = 777; + int i, rank, size, err, count; + MPI_Aint lb, sendextent, recvextent; + MPI_Request* requests; + + // FIXME: check implementation + rank = smpi_comm_rank(comm); + size = smpi_comm_size(comm); + DEBUG1("<%d> algorithm alltoall_bruck() called.", rank); + err = smpi_datatype_extent(sendtype, &lb, &sendextent); + err = smpi_datatype_extent(recvtype, &lb, &recvextent); + /* Local copy from self */ + err = smpi_datatype_copy(&((char*)sendbuf)[rank * sendextent], sendcount, sendtype, &((char*)recvbuf)[rank * recvextent], recvcount, recvtype); + if(err == MPI_SUCCESS && size > 1) { + /* Initiate all send/recv to/from others. */ + requests = xbt_new(MPI_Request, 2 * (size - 1)); + count = 0; + /* Create all receives that will be posted first */ + for(i = 0; i < size; ++i) { + if(i == rank) { + DEBUG3("<%d> skip request creation [src = %d, recvcount = %d]", rank, i, recvcount); + continue; + } + requests[count] = smpi_irecv_init(&((char*)recvbuf)[i * recvextent], recvcount, recvtype, i, system_tag, comm); + count++; + } + /* Now create all sends */ + for(i = 0; i < size; ++i) { + if(i == rank) { + DEBUG3("<%d> skip request creation [dst = %d, sendcount = %d]", rank, i, sendcount); + continue; + } + requests[count] = smpi_isend_init(&((char*)sendbuf)[i * sendextent], sendcount, sendtype, i, system_tag, comm); + count++; + } + /* Wait for them all.*/ + smpi_mpi_startall(count, requests); + DEBUG2("<%d> wait for %d requests", rank, count); + smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE); + xbt_free(requests); + } return MPI_SUCCESS; } @@ -211,25 +251,20 @@ int smpi_coll_tuned_alltoall_basic_linear(void *sendbuf, int sendcount, MPI_Data /* Post all receives first -- a simple optimization */ count = 0; for(i = (rank + 1) % size; i != rank; i = (i + 1) % size) { - requests[count] = smpi_mpi_irecv(&((char*)recvbuf)[i * recvinc], recvcount, recvtype, i, system_tag, comm); + requests[count] = smpi_irecv_init(&((char*)recvbuf)[i * recvinc], recvcount, recvtype, i, system_tag, comm); count++; } - /* Now post all sends in reverse order + /* Now post all sends in reverse order * - We would like to minimize the search time through message queue * when messages actually arrive in the order in which they were posted. * TODO: check the previous assertion */ for(i = (rank + size - 1) % size; i != rank; i = (i + size - 1) % size ) { - requests[count] = smpi_mpi_isend(&((char*)sendbuf)[i * sendinc], sendcount, sendtype, i, system_tag, comm); + requests[count] = smpi_isend_init(&((char*)sendbuf)[i * sendinc], sendcount, sendtype, i, system_tag, comm); count++; } - /* Wait for them all. If there's an error, note that we don't - * care what the error was -- just that there *was* an error. The - * PML will finish all requests, even if one or more of them fail. - * i.e., by the end of this call, all the requests are free-able. - * So free them anyway -- even if there was an error, and return - * the error after we free everything. - */ + /* Wait for them all.*/ + smpi_mpi_startall(count, requests); DEBUG2("<%d> wait for %d requests", rank, count); smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE); xbt_free(requests); @@ -242,9 +277,9 @@ int smpi_coll_tuned_alltoall_basic_linear(void *sendbuf, int sendcount, MPI_Data * * this algorithm performs size steps (1<=s<=size) and * at each step s, a process p sends iand receive to.from a unique distinct remote process - * size=5 : s=1: 4->0->1, 0->1->2, 1->2->3, ... + * size=5 : s=1: 4->0->1, 0->1->2, 1->2->3, ... * s=2: 3->0->2, 4->1->3, 0->2->4, 1->3->0 , 2->4->1 - * .... + * .... * Openmpi calls this routine when the message size sent to each rank is greater than 3000 bytes **/ int smpi_coll_tuned_alltoall_pairwise(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) { @@ -291,7 +326,7 @@ int smpi_coll_basic_alltoallv(void* sendbuf, int* sendcounts, int* senddisps, MP DEBUG3("<%d> skip request creation [src = %d, recvcounts[src] = %d]", rank, i, recvcounts[i]); continue; } - requests[count] = smpi_mpi_irecv(&((char*)recvbuf)[recvdisps[i] * recvextent], recvcounts[i], recvtype, i, system_tag, comm); + requests[count] = smpi_irecv_init(&((char*)recvbuf)[recvdisps[i] * recvextent], recvcounts[i], recvtype, i, system_tag, comm); count++; } /* Now create all sends */ @@ -300,16 +335,11 @@ int smpi_coll_basic_alltoallv(void* sendbuf, int* sendcounts, int* senddisps, MP DEBUG3("<%d> skip request creation [dst = %d, sendcounts[dst] = %d]", rank, i, sendcounts[i]); continue; } - requests[count] = smpi_mpi_isend(&((char*)sendbuf)[senddisps[i] * sendextent], sendcounts[i], sendtype, i, system_tag, comm); + requests[count] = smpi_isend_init(&((char*)sendbuf)[senddisps[i] * sendextent], sendcounts[i], sendtype, i, system_tag, comm); count++; } - /* Wait for them all. If there's an error, note that we don't - * care what the error was -- just that there *was* an error. The - * PML will finish all requests, even if one or more of them fail. - * i.e., by the end of this call, all the requests are free-able. - * So free them anyway -- even if there was an error, and return - * the error after we free everything. - */ + /* Wait for them all.*/ + smpi_mpi_startall(count, requests); DEBUG2("<%d> wait for %d requests", rank, count); smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE); xbt_free(requests);