-/* $Id$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 <stdio.h>
-#include <stdlib.h>
#include <string.h>
+#include <assert.h>
#include "private.h"
-
-/* proctree taken and translated from P2P-MPI */
-
-struct proc_tree {
- int PROCTREE_A;
- int numChildren;
- int * child;
- int parent;
- int me;
- int root;
- int isRoot;
+XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_coll, smpi,
+ "Logging specific to SMPI (coll)");
+
+struct s_proc_tree {
+ int PROCTREE_A;
+ int numChildren;
+ int *child;
+ int parent;
+ int me;
+ int root;
+ int isRoot;
};
-typedef struct proc_tree * proc_tree_t;
-
-
-
-/* prototypes */
-void build_tree( int index, int extent, proc_tree_t *tree);
-proc_tree_t alloc_tree( int arity );
-void free_tree( proc_tree_t tree);
-void print_tree(proc_tree_t tree);
-int binomial_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm);
-
-
+typedef struct s_proc_tree *proc_tree_t;
/**
* alloc and init
**/
-proc_tree_t alloc_tree( int arity ) {
- proc_tree_t tree = malloc(1*sizeof(struct proc_tree));
- int i;
-
- tree->PROCTREE_A = arity;
- tree->isRoot = 0;
- tree->numChildren = 0;
- tree->child = malloc(arity*sizeof(int));
- for (i=0; i < arity; i++) {
- tree->child[i] = -1;
- }
- tree->root = -1;
- tree->parent = -1;
- return( tree );
+static proc_tree_t alloc_tree(int arity)
+{
+ proc_tree_t tree;
+ int i;
+
+ tree = xbt_new(struct s_proc_tree, 1);
+ tree->PROCTREE_A = arity;
+ tree->isRoot = 0;
+ tree->numChildren = 0;
+ tree->child = xbt_new(int, arity);
+ for (i = 0; i < arity; i++) {
+ tree->child[i] = -1;
+ }
+ tree->root = -1;
+ tree->parent = -1;
+ return tree;
}
/**
* free
**/
-void free_tree( proc_tree_t tree) {
- free (tree->child );
- free(tree);
+static void free_tree(proc_tree_t tree)
+{
+ xbt_free(tree->child);
+ xbt_free(tree);
}
-
-
/**
* Build the tree depending on a process rank (index) and the group size (extent)
* @param index the rank of the calling process
* @param extent the total number of processes
**/
-void build_tree( int index, int extent, proc_tree_t *tree) {
- int places = (*tree)->PROCTREE_A * index;
- int i;
- int ch;
- int pr;
-
- (*tree)->me = index;
- (*tree)->root = 0 ;
-
- for (i = 1; i <= (*tree)->PROCTREE_A; i++) {
- ++places;
- ch = (*tree)->PROCTREE_A * index + i + (*tree)->root;
- //printf("places %d\n",places);
- ch %= extent;
- if (places < extent) {
- //printf("ch <%d> = <%d>\n",i,ch);
- //printf("adding to the tree at index <%d>\n\n",i-1);
- (*tree)->child[i - 1] = ch;
- (*tree)->numChildren++;
- }
- else {
- //fprintf(stderr,"not adding to the tree\n");
- }
- }
- //fprintf(stderr,"procTree.numChildren <%d>\n",(*tree)->numChildren);
-
- if (index == (*tree)->root) {
- (*tree)->isRoot = 1;
- }
- else {
- (*tree)->isRoot = 0;
- pr = (index - 1) / (*tree)->PROCTREE_A;
- (*tree)->parent = pr;
- }
+static void build_tree(int index, int extent, proc_tree_t * tree)
+{
+ int places = (*tree)->PROCTREE_A * index;
+ int i, ch, pr;
+
+ (*tree)->me = index;
+ (*tree)->root = 0;
+ for (i = 1; i <= (*tree)->PROCTREE_A; i++) {
+ ++places;
+ ch = (*tree)->PROCTREE_A * index + i + (*tree)->root;
+ ch %= extent;
+ if (places < extent) {
+ (*tree)->child[i - 1] = ch;
+ (*tree)->numChildren++;
+ }
+ }
+ if (index == (*tree)->root) {
+ (*tree)->isRoot = 1;
+ } else {
+ (*tree)->isRoot = 0;
+ pr = (index - 1) / (*tree)->PROCTREE_A;
+ (*tree)->parent = pr;
+ }
}
+/**
+ * bcast
+ **/
+static void tree_bcast(void *buf, int count, MPI_Datatype datatype,
+ int root, MPI_Comm comm, proc_tree_t tree)
+{
+ int system_tag = 999; // used negative int but smpi_create_request() declares this illegal (to be checked)
+ int rank, i;
+ MPI_Request *requests;
+
+ rank = smpi_comm_rank(comm);
+ /* wait for data from my parent in the tree */
+ if (!tree->isRoot) {
+ DEBUG3("<%d> tree_bcast(): i am not root: recv from %d, tag=%d)",
+ rank, tree->parent, system_tag + rank);
+ 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)", 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_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);
+}
-void print_tree(proc_tree_t tree) {
- int i;
- char *spacer;
- if (-1 != tree->parent ) {
- printf("[%d]\n +---",tree->parent);
- spacer= strdup(" ");
- }
- else {
- spacer=strdup("");
- }
- printf("<%d>\n",tree->me);
- for (i=0;i < tree->numChildren; i++) {
- printf("%s +--- %d\n", spacer,tree->child[i]);
- }
- free(spacer);
+/**
+ * anti-bcast
+ **/
+static void tree_antibcast(void *buf, int count, MPI_Datatype datatype,
+ int root, MPI_Comm comm, proc_tree_t tree)
+{
+ int system_tag = 999; // used negative int but smpi_create_request() declares this illegal (to be checked)
+ int rank, i;
+ MPI_Request *requests;
+
+ rank = smpi_comm_rank(comm);
+ // everyone sends to its parent, except root.
+ if (!tree->isRoot) {
+ DEBUG3("<%d> tree_antibcast(): i am not root: send to %d, tag=%d)",
+ rank, tree->parent, system_tag + rank);
+ smpi_mpi_send(buf, count, datatype, tree->parent, system_tag + rank,
+ comm);
+ }
+ //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)", 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_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
+ * 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);
+ int rank, size;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ build_tree(rank, size, &tree);
+ tree_bcast(buf, count, datatype, root, comm, tree);
+ free_tree(tree);
+}
+
+/**
+ * barrier with a binary, ternary, or whatever tree ..
**/
-int binomial_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root,
- MPI_Comm comm)
+void nary_tree_barrier(MPI_Comm comm, int arity)
{
- int system_tag = 999; // used negative int but smpi_create_request() declares this illegal (to be checked)
- int rank;
- int retval = MPI_SUCCESS;
- int i;
- smpi_mpi_request_t request;
- smpi_mpi_request_t * requests;
- void **tmpbufs;
-
- rank = smpi_mpi_comm_rank(comm);
- proc_tree_t tree = alloc_tree( 2 ); // arity=2: a binomial tree
-
- build_tree( rank, comm->size, &tree );
- /* wait for data from my parent in the tree */
- if (!tree->isRoot) {
-#ifdef DEBUG_STEPH
- printf("[%d] recv(%d from %d, tag=%d)\n",rank,rank, tree->parent, system_tag+rank);
-#endif
- retval = smpi_create_request(buf, count, datatype,
- tree->parent, rank,
- system_tag + rank,
- comm, &request);
- if (MPI_SUCCESS != retval) {
- printf("** internal error: smpi_create_request() rank=%d returned retval=%d, %s:%d\n",
- rank,retval,__FILE__,__LINE__);
- }
- smpi_mpi_irecv(request);
-#ifdef DEBUG_STEPH
- printf("[%d] waiting on irecv from %d\n",rank , tree->parent);
-#endif
- smpi_mpi_wait(request, MPI_STATUS_IGNORE);
- }
-
- tmpbufs = xbt_malloc( tree->numChildren * sizeof(void *));
- requests = xbt_malloc( tree->numChildren * sizeof(smpi_mpi_request_t));
-#ifdef DEBUG_STEPH
- printf("[%d] creates %d requests\n",rank,tree->numChildren);
-#endif
-
- /* iniates sends to ranks lower in the tree */
- for (i=0; i < tree->numChildren; i++) {
- if (tree->child[i] != -1) {
-#ifdef DEBUG_STEPH
- printf("[%d] send(%d->%d, tag=%d)\n",rank,rank, tree->child[i], system_tag+tree->child[i]);
-#endif
- tmpbufs[i] = xbt_malloc( count * datatype->size);
- memcpy( tmpbufs[i], buf, count * datatype->size * sizeof(char));
- retval = smpi_create_request(tmpbufs[i], count, datatype,
- rank, tree->child[i],
- system_tag + tree->child[i],
- comm, &(requests[i]));
-#ifdef DEBUG_STEPH
- printf("[%d] after create req[%d]=%p req->(src=%d,dst=%d)\n",rank , i, requests[i],requests[i]->src,requests[i]->dst );
-#endif
- if (MPI_SUCCESS != retval) {
- printf("** internal error: smpi_create_request() rank=%d returned retval=%d, %s:%d\n",
- rank,retval,__FILE__,__LINE__);
- }
- smpi_mpi_isend(requests[i]);
- /* FIXME : we should not wait immediately here. See next FIXME. */
- smpi_mpi_wait( requests[i], MPI_STATUS_IGNORE);
- xbt_free(tmpbufs[i]);
- xbt_free(requests[i]);
- }
- }
- /* FIXME : normally, we sould wait only once all isend have been issued:
- * this is the following commented code. It deadlocks, probably because
- * of a bug in the sender process */
-
- /* wait for completion of sends */
- /* printf("[%d] wait for %d send completions\n",rank,tree->numChildren);
- smpi_mpi_waitall( tree->numChildren, requests, MPI_STATUS_IGNORE);
- printf("[%d] reqs completed\n)",rank);
- */
-
- xbt_free(tmpbufs);
- xbt_free(requests);
- return(retval);
+ proc_tree_t tree = alloc_tree(arity);
+ int rank, size;
+ char dummy = '$';
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ build_tree(rank, size, &tree);
+ tree_antibcast(&dummy, 1, MPI_CHAR, 0, comm, tree);
+ tree_bcast(&dummy, 1, MPI_CHAR, 0, comm, tree);
+ free_tree(tree);
}
/**
- * example usage
+ * Alltoall Bruck
+ *
+ * Openmpi calls this routine when the message size sent to each rank < 2000 bytes and size < 12
**/
-/*
- * int main() {
-
- int rank;
- int size=12;
-
- proc_tree_t tree;
- for (rank=0;rank<size;rank++) {
- printf("--------------tree for rank %d ----------\n",rank);
- tree = alloc_tree( 2 );
- build_tree( rank, size, &tree );
- print_tree( tree );
- free_tree( tree );
-
+int smpi_coll_tuned_alltoall_bruck(void *sendbuf, int sendcount,
+ MPI_Datatype sendtype, void *recvbuf,
+ int recvcount, MPI_Datatype recvtype,
+ MPI_Comm comm)
+{
+ int system_tag = 777;
+ int i, rank, size, err, count;
+ MPI_Aint lb;
+ MPI_Aint sendextent = 0;
+ MPI_Aint recvextent = 0;
+ 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;
}
- printf("-------------- bcast ----------\n");
- for (rank=0;rank<size;rank++) {
- bcast( rank, size );
+ 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;
+}
-
+/**
+ * Alltoall basic_linear
+ **/
+int smpi_coll_tuned_alltoall_basic_linear(void *sendbuf, int sendcount,
+ MPI_Datatype sendtype,
+ void *recvbuf, int recvcount,
+ MPI_Datatype recvtype,
+ MPI_Comm comm)
+{
+ int system_tag = 888;
+ int i, rank, size, err, count;
+ MPI_Aint lb;
+ MPI_Aint sendinc = 0;
+ MPI_Aint recvinc = 0;
+ MPI_Request *requests;
+
+ /* Initialize. */
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ DEBUG1("<%d> algorithm alltoall_basic_linear() called.", rank);
+ err = smpi_datatype_extent(sendtype, &lb, &sendinc);
+ err = smpi_datatype_extent(recvtype, &lb, &recvinc);
+ sendinc *= sendcount;
+ recvinc *= recvcount;
+ /* simple optimization */
+ err =
+ smpi_datatype_copy(&((char *) sendbuf)[rank * sendinc], sendcount,
+ sendtype, &((char *) recvbuf)[rank * recvinc],
+ recvcount, recvtype);
+ if (err == MPI_SUCCESS && size > 1) {
+ /* Initiate all send/recv to/from others. */
+ requests = xbt_new(MPI_Request, 2 * (size - 1));
+ /* Post all receives first -- a simple optimization */
+ count = 0;
+ for (i = (rank + 1) % size; i != rank; i = (i + 1) % size) {
+ requests[count] =
+ smpi_irecv_init(&((char *) recvbuf)[i * recvinc], recvcount,
+ recvtype, i, system_tag, comm);
+ count++;
+ }
+ /* 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_isend_init(&((char *) sendbuf)[i * sendinc], 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 err;
}
-*/
-
+/**
+ * Alltoall pairwise
+ *
+ * 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, ...
+ * 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)
+{
+ int system_tag = 999;
+ int rank, size, step, sendto, recvfrom, sendsize, recvsize;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ DEBUG1("<%d> algorithm alltoall_pairwise() called.", rank);
+ sendsize = smpi_datatype_size(sendtype);
+ recvsize = smpi_datatype_size(recvtype);
+ /* Perform pairwise exchange - starting from 1 so the local copy is last */
+ for (step = 1; step < size + 1; step++) {
+ /* who do we talk to in this step? */
+ sendto = (rank + step) % size;
+ recvfrom = (rank + size - step) % size;
+ /* send and receive */
+ smpi_mpi_sendrecv(&((char *) sendbuf)[sendto * sendsize * sendcount],
+ sendcount, sendtype, sendto, system_tag,
+ &((char *) recvbuf)[recvfrom * recvsize * recvcount],
+ recvcount, recvtype, recvfrom, system_tag, comm,
+ MPI_STATUS_IGNORE);
+ }
+ return MPI_SUCCESS;
+}
+int smpi_coll_basic_alltoallv(void *sendbuf, int *sendcounts,
+ int *senddisps, MPI_Datatype sendtype,
+ void *recvbuf, int *recvcounts,
+ int *recvdisps, MPI_Datatype recvtype,
+ MPI_Comm comm)
+{
+ int system_tag = 889;
+ int i, rank, size, err, count;
+ MPI_Aint lb;
+ MPI_Aint sendextent = 0;
+ MPI_Aint recvextent = 0;
+ MPI_Request *requests;
+
+ /* Initialize. */
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ DEBUG1("<%d> algorithm basic_alltoallv() 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)[senddisps[rank] * sendextent],
+ sendcounts[rank], sendtype,
+ &((char *) recvbuf)[recvdisps[rank] * recvextent],
+ recvcounts[rank], 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 || recvcounts[i] == 0) {
+ DEBUG3
+ ("<%d> skip request creation [src = %d, recvcounts[src] = %d]",
+ rank, i, recvcounts[i]);
+ continue;
+ }
+ requests[count] =
+ smpi_irecv_init(&((char *) recvbuf)[recvdisps[i] * recvextent],
+ recvcounts[i], recvtype, i, system_tag, comm);
+ count++;
+ }
+ /* Now create all sends */
+ for (i = 0; i < size; ++i) {
+ if (i == rank || sendcounts[i] == 0) {
+ DEBUG3
+ ("<%d> skip request creation [dst = %d, sendcounts[dst] = %d]",
+ rank, i, sendcounts[i]);
+ continue;
+ }
+ requests[count] =
+ smpi_isend_init(&((char *) sendbuf)[senddisps[i] * sendextent],
+ sendcounts[i], 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 err;
+}