-/* $Id$tag */
-
/* smpi_coll.c -- various optimized routing for collectives */
-/* Copyright (c) 2009 Stephane Genaud. */
-/* All rights reserved. */
+/* Copyright (c) 2009-2015. 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"
+#include "colls/colls.h"
+#include "simgrid/sg_config.h"
+
+XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_coll, smpi,
+ "Logging specific to SMPI (coll)");
+
+s_mpi_coll_description_t mpi_coll_gather_description[] = {
+ {"default",
+ "gather default collective",
+ smpi_mpi_gather},
+COLL_GATHERS(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
-/* proctree taken and translated from P2P-MPI */
+s_mpi_coll_description_t mpi_coll_allgather_description[] = {
+ {"default",
+ "allgather default collective",
+ smpi_mpi_allgather},
+COLL_ALLGATHERS(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
-struct proc_tree {
- int PROCTREE_A;
- int numChildren;
- int * child;
- int parent;
- int me;
- int root;
- int isRoot;
+s_mpi_coll_description_t mpi_coll_allgatherv_description[] = {
+ {"default",
+ "allgatherv default collective",
+ smpi_mpi_allgatherv},
+COLL_ALLGATHERVS(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
};
-typedef struct proc_tree * proc_tree_t;
+s_mpi_coll_description_t mpi_coll_allreduce_description[] = {
+ {"default",
+ "allreduce default collective",
+ smpi_mpi_allreduce},
+COLL_ALLREDUCES(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
+s_mpi_coll_description_t mpi_coll_reduce_scatter_description[] = {
+ {"default",
+ "reduce_scatter default collective",
+ smpi_mpi_reduce_scatter},
+COLL_REDUCE_SCATTERS(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
-/* 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);
+s_mpi_coll_description_t mpi_coll_scatter_description[] = {
+ {"default",
+ "scatter default collective",
+ smpi_mpi_scatter},
+COLL_SCATTERS(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
+s_mpi_coll_description_t mpi_coll_barrier_description[] = {
+ {"default",
+ "barrier default collective - ompi selector",
+ smpi_coll_tuned_barrier_ompi},
+COLL_BARRIERS(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
+s_mpi_coll_description_t mpi_coll_alltoall_description[] = {
+ {"default",
+ "Ompi alltoall default collective",
+ smpi_coll_tuned_alltoall_ompi2},
+COLL_ALLTOALLS(COLL_DESCRIPTION, COLL_COMMA),
+ {"bruck",
+ "Alltoall Bruck (SG) collective",
+ smpi_coll_tuned_alltoall_bruck},
+ {"basic_linear",
+ "Alltoall basic linear (SG) collective",
+ smpi_coll_tuned_alltoall_basic_linear},
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
+s_mpi_coll_description_t mpi_coll_alltoallv_description[] = {
+ {"default",
+ "Ompi alltoallv default collective",
+ smpi_coll_basic_alltoallv},
+COLL_ALLTOALLVS(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
-/**
- * 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 );
-}
+s_mpi_coll_description_t mpi_coll_bcast_description[] = {
+ {"default",
+ "bcast default collective - ompi selector",
+ smpi_coll_tuned_bcast_ompi},
+COLL_BCASTS(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
-/**
- * free
- **/
-void free_tree( proc_tree_t tree) {
- free (tree->child );
- free(tree);
-}
+s_mpi_coll_description_t mpi_coll_reduce_description[] = {
+ {"default",
+ "reduce default collective",
+ smpi_mpi_reduce},
+COLL_REDUCES(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
-/**
- * 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;
- }
+/** Displays the long description of all registered models, and quit */
+void coll_help(const char *category, s_mpi_coll_description_t * table)
+{
+ int i;
+ printf("Long description of the %s models accepted by this simulator:\n",
+ category);
+ for (i = 0; table[i].name; i++)
+ printf(" %s: %s\n", table[i].name, table[i].description);
}
+int find_coll_description(s_mpi_coll_description_t * table,
+ char *name, const char *desc)
+{
+ int i;
+ char *name_list = NULL;
+ int selector_on=0;
+ if(name==NULL){//no argument provided, use active selector's algorithm
+ name=(char*)sg_cfg_get_string("smpi/coll_selector");
+ selector_on=1;
+ }
+ for (i = 0; table[i].name; i++)
+ if (!strcmp(name, table[i].name)) {
+ if (strcmp(table[i].name,"default"))
+ XBT_INFO("Switch to algorithm %s for collective %s",table[i].name,desc);
+ return i;
+ }
+
+ if(selector_on){
+ // collective seems not handled by the active selector, try with default one
+ name=(char*)"default";
+ for (i = 0; table[i].name; i++)
+ if (!strcmp(name, table[i].name)) {
+ return i;
+ }
+ }
+ if (!table[0].name)
+ xbt_die("No collective is valid for '%s'! This is a bug.",name);
+ name_list = xbt_strdup(table[0].name);
+ for (i = 1; table[i].name; i++) {
+ name_list =
+ xbt_realloc(name_list,
+ strlen(name_list) + strlen(table[i].name) + 3);
+ strcat(name_list, ", ");
+ strcat(name_list, table[i].name);
+ }
+ xbt_die("Collective '%s' is invalid! Valid collectives are: %s.", name, name_list);
+ return -1;
+}
-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);
+int (*mpi_coll_gather_fun)(void *, int, MPI_Datatype, void*, int, MPI_Datatype, int root, MPI_Comm);
+int (*mpi_coll_allgather_fun)(void *, int, MPI_Datatype, void*, int, MPI_Datatype, MPI_Comm);
+int (*mpi_coll_allgatherv_fun)(void *, int, MPI_Datatype, void*, int*, int*, MPI_Datatype, MPI_Comm);
+int (*mpi_coll_allreduce_fun)(void *sbuf, void *rbuf, int rcount, MPI_Datatype dtype, MPI_Op op, MPI_Comm comm);
+int (*mpi_coll_alltoall_fun)(void *, int, MPI_Datatype, void*, int, MPI_Datatype, MPI_Comm);
+int (*mpi_coll_alltoallv_fun)(void *, int*, int*, MPI_Datatype, void*, int*, int*, MPI_Datatype, MPI_Comm);
+int (*mpi_coll_bcast_fun)(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm com);
+int (*mpi_coll_reduce_fun)(void *buf, void *rbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm);
+int (*mpi_coll_reduce_scatter_fun)(void *sbuf, void *rbuf, int *rcounts,MPI_Datatype dtype,MPI_Op op,MPI_Comm comm);
+int (*mpi_coll_scatter_fun)(void *sendbuf, int sendcount, MPI_Datatype sendtype,void *recvbuf, int recvcount, MPI_Datatype recvtype,int root, MPI_Comm comm);
+int (*mpi_coll_barrier_fun)(MPI_Comm comm);
+
+
+int smpi_coll_tuned_alltoall_ompi2(void *sendbuf, int sendcount,
+ MPI_Datatype sendtype, void *recvbuf,
+ int recvcount, MPI_Datatype recvtype,
+ MPI_Comm comm)
+{
+ int size, sendsize;
+ size = smpi_comm_size(comm);
+ sendsize = smpi_datatype_size(sendtype) * sendcount;
+ if (sendsize < 200 && size > 12) {
+ return
+ smpi_coll_tuned_alltoall_bruck(sendbuf, sendcount, sendtype,
+ recvbuf, recvcount, recvtype,
+ comm);
+ } else if (sendsize < 3000) {
+ return
+ smpi_coll_tuned_alltoall_basic_linear(sendbuf, sendcount,
+ sendtype, recvbuf,
+ recvcount, recvtype, comm);
+ } else {
+ return
+ smpi_coll_tuned_alltoall_ring(sendbuf, sendcount, sendtype,
+ recvbuf, recvcount, recvtype,
+ comm);
+ }
}
-
+
/**
- * bcast
+ * Alltoall Bruck
+ *
+ * Openmpi calls this routine when the message size sent to each rank < 2000 bytes and size < 12
+ * FIXME: uh, check smpi_pmpi again, but this routine is called for > 12, not
+ * less...
**/
-int binomial_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root,
- MPI_Comm comm)
+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 = 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);
+ int system_tag = 777;
+ int i, rank, size, err, count;
+ MPI_Aint lb;
+ MPI_Aint sendext = 0;
+ MPI_Aint recvext = 0;
+ MPI_Request *requests;
+
+ // FIXME: check implementation
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ XBT_DEBUG("<%d> algorithm alltoall_bruck() called.", rank);
+ smpi_datatype_extent(sendtype, &lb, &sendext);
+ smpi_datatype_extent(recvtype, &lb, &recvext);
+ /* Local copy from self */
+ err =
+ smpi_datatype_copy((char *)sendbuf + rank * sendcount * sendext,
+ sendcount, sendtype,
+ (char *)recvbuf + rank * recvcount * recvext,
+ 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) {
+ XBT_DEBUG("<%d> skip request creation [src = %d, recvcount = %d]",
+ rank, i, recvcount);
+ continue;
+ }
+ requests[count] =
+ smpi_irecv_init((char *)recvbuf + i * recvcount * recvext, recvcount,
+ recvtype, i, system_tag, comm);
+ count++;
+ }
+ /* Now create all sends */
+ for (i = 0; i < size; ++i) {
+ if (i == rank) {
+ XBT_DEBUG("<%d> skip request creation [dst = %d, sendcount = %d]",
+ rank, i, sendcount);
+ continue;
+ }
+ requests[count] =
+ smpi_isend_init((char *)sendbuf + i * sendcount * sendext, sendcount,
+ sendtype, i, system_tag, comm);
+ count++;
+ }
+ /* Wait for them all. */
+ smpi_mpi_startall(count, requests);
+ XBT_DEBUG("<%d> wait for %d requests", rank, count);
+ smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE);
+ for(i = 0; i < count; i++) {
+ if(requests[i]!=MPI_REQUEST_NULL) smpi_mpi_request_free(&requests[i]);
+ }
+ xbt_free(requests);
+ }
+ return MPI_SUCCESS;
}
/**
- * example usage
+ * Alltoall basic_linear (STARMPI:alltoall-simple)
**/
-/*
- * 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_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 = 0, sendext = 0, recvext = 0;
+ MPI_Request *requests;
+
+ /* Initialize. */
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ XBT_DEBUG("<%d> algorithm alltoall_basic_linear() called.", rank);
+ smpi_datatype_extent(sendtype, &lb, &sendext);
+ smpi_datatype_extent(recvtype, &lb, &recvext);
+ /* simple optimization */
+ err = smpi_datatype_copy((char *)sendbuf + rank * sendcount * sendext,
+ sendcount, sendtype,
+ (char *)recvbuf + rank * recvcount * recvext,
+ 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 * recvcount * recvext, 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 * sendcount * sendext, sendcount,
+ sendtype, i, system_tag, comm);
+ count++;
+ }
+ /* Wait for them all. */
+ smpi_mpi_startall(count, requests);
+ XBT_DEBUG("<%d> wait for %d requests", rank, count);
+ smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE);
+ for(i = 0; i < count; i++) {
+ if(requests[i]!=MPI_REQUEST_NULL) smpi_mpi_request_free(&requests[i]);
+ }
+ xbt_free(requests);
+ }
+ return err;
+}
+
+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 = 0, sendext = 0, recvext = 0;
+ MPI_Request *requests;
+
+ /* Initialize. */
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ XBT_DEBUG("<%d> algorithm basic_alltoallv() called.", rank);
+ smpi_datatype_extent(sendtype, &lb, &sendext);
+ smpi_datatype_extent(recvtype, &lb, &recvext);
+ /* Local copy from self */
+ err =
+ smpi_datatype_copy((char *)sendbuf + senddisps[rank] * sendext,
+ sendcounts[rank], sendtype,
+ (char *)recvbuf + recvdisps[rank] * recvext,
+ 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) {
+ XBT_DEBUG
+ ("<%d> skip request creation [src = %d, recvcounts[src] = %d]",
+ rank, i, recvcounts[i]);
+ continue;
}
- printf("-------------- bcast ----------\n");
- for (rank=0;rank<size;rank++) {
- bcast( rank, size );
+ requests[count] =
+ smpi_irecv_init((char *)recvbuf + recvdisps[i] * recvext,
+ recvcounts[i], recvtype, i, system_tag, comm);
+ count++;
+ }
+ /* Now create all sends */
+ for (i = 0; i < size; ++i) {
+ if (i == rank || sendcounts[i] == 0) {
+ XBT_DEBUG
+ ("<%d> skip request creation [dst = %d, sendcounts[dst] = %d]",
+ rank, i, sendcounts[i]);
+ continue;
}
-
-
+ requests[count] =
+ smpi_isend_init((char *)sendbuf + senddisps[i] * sendext,
+ sendcounts[i], sendtype, i, system_tag, comm);
+ count++;
+ }
+ /* Wait for them all. */
+ smpi_mpi_startall(count, requests);
+ XBT_DEBUG("<%d> wait for %d requests", rank, count);
+ smpi_mpi_waitall(count, requests, MPI_STATUS_IGNORE);
+ for(i = 0; i < count; i++) {
+ if(requests[i]!=MPI_REQUEST_NULL) smpi_mpi_request_free(&requests[i]);
+ }
+ xbt_free(requests);
+ }
+ return err;
}
-*/
-
-
-
