/* smpi_coll.c -- various optimized routing for collectives */
-/* Copyright (c) 2009, 2010. The SimGrid Team.
+/* Copyright (c) 2009-2013. The SimGrid Team.
* All rights reserved. */
/* This program is free software; you can redistribute it and/or modify it
#include <assert.h>
#include "private.h"
+#include "colls/colls.h"
+#include "simgrid/sg_config.h"
+
+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 */
+};
+
+
+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 */
+};
+
+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 */
+};
+
+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 */
+};
+
+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",
+ smpi_mpi_barrier},
+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 */
+};
+
+s_mpi_coll_description_t mpi_coll_bcast_description[] = {
+ {"default",
+ "bcast default collective",
+ smpi_mpi_bcast},
+COLL_BCASTS(COLL_DESCRIPTION, COLL_COMMA),
+ {NULL, NULL, NULL} /* this array must be NULL terminated */
+};
+
+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 */
+};
+
+
+
+/** 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)
+{
+ 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)) {
+ 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! This is a bug.");
+ 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;
+}
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_coll, smpi,
"Logging specific to SMPI (coll)");
+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);
struct s_proc_tree {
int PROCTREE_A;
int numChildren;
* bcast
**/
static void tree_bcast(void *buf, int count, MPI_Datatype datatype,
- int root, MPI_Comm comm, proc_tree_t tree)
+ 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 system_tag = COLL_TAG_BCAST;
int rank, i;
MPI_Request *requests;
* anti-bcast
**/
static void tree_antibcast(void *buf, int count, MPI_Datatype datatype,
- int root, MPI_Comm comm, proc_tree_t tree)
+ 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 system_tag = COLL_TAG_BCAST;
int rank, i;
MPI_Request *requests;
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
build_tree(root, rank, size, &tree);
- tree_bcast(buf, count, datatype, root, comm, tree);
+ tree_bcast(buf, count, datatype, comm, tree);
free_tree(tree);
}
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
build_tree(0, rank, size, &tree);
- tree_antibcast(&dummy, 1, MPI_CHAR, 0, comm, tree);
- tree_bcast(&dummy, 1, MPI_CHAR, 0, comm, tree);
+ tree_antibcast(&dummy, 1, MPI_CHAR, comm, tree);
+ tree_bcast(&dummy, 1, MPI_CHAR, comm, tree);
free_tree(tree);
}
+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);
+ }
+}
+
/**
* 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 smpi_coll_tuned_alltoall_bruck(void *sendbuf, int sendcount,
MPI_Datatype sendtype, void *recvbuf,
int system_tag = 777;
int i, rank, size, err, count;
MPI_Aint lb;
- MPI_Aint sendextent = 0;
- MPI_Aint recvextent = 0;
+ 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);
- err = smpi_datatype_extent(sendtype, &lb, &sendextent);
- err = smpi_datatype_extent(recvtype, &lb, &recvextent);
+ err = smpi_datatype_extent(sendtype, &lb, &sendext);
+ err = smpi_datatype_extent(recvtype, &lb, &recvext);
/* Local copy from self */
err =
- smpi_datatype_copy(&((char *) sendbuf)[rank * sendextent], sendcount,
- sendtype, &((char *) recvbuf)[rank * recvextent],
+ 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. */
continue;
}
requests[count] =
- smpi_irecv_init(&((char *) recvbuf)[i * recvextent], recvcount,
+ smpi_irecv_init((char *)recvbuf + i * recvcount * recvext, recvcount,
recvtype, i, system_tag, comm);
count++;
}
continue;
}
requests[count] =
- smpi_isend_init(&((char *) sendbuf)[i * sendextent], sendcount,
+ smpi_isend_init((char *)sendbuf + i * sendcount * sendext, sendcount,
sendtype, i, system_tag, comm);
count++;
}
}
/**
- * Alltoall basic_linear
+ * Alltoall basic_linear (STARMPI:alltoall-simple)
**/
int smpi_coll_tuned_alltoall_basic_linear(void *sendbuf, int sendcount,
MPI_Datatype sendtype,
{
int system_tag = 888;
int i, rank, size, err, count;
- MPI_Aint lb;
- MPI_Aint sendinc = 0;
- MPI_Aint recvinc = 0;
+ 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);
- err = smpi_datatype_extent(sendtype, &lb, &sendinc);
- err = smpi_datatype_extent(recvtype, &lb, &recvinc);
- sendinc *= sendcount;
- recvinc *= recvcount;
+ err = smpi_datatype_extent(sendtype, &lb, &sendext);
+ err = smpi_datatype_extent(recvtype, &lb, &recvext);
/* simple optimization */
- err =
- smpi_datatype_copy(&((char *) sendbuf)[rank * sendinc], sendcount,
- sendtype, &((char *) recvbuf)[rank * recvinc],
- recvcount, recvtype);
+ 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;
for (i = (rank + 1) % size; i != rank; i = (i + 1) % size) {
requests[count] =
- smpi_irecv_init(&((char *) recvbuf)[i * recvinc], recvcount,
+ smpi_irecv_init((char *)recvbuf + i * recvcount * recvext, recvcount,
recvtype, i, system_tag, comm);
count++;
}
* 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) {
+ for (i = (rank + size - 1) % size; i != rank; i = (i + size - 1) % size) {
requests[count] =
- smpi_isend_init(&((char *) sendbuf)[i * sendinc], sendcount,
+ smpi_isend_init((char *)sendbuf + i * sendcount * sendext, sendcount,
sendtype, i, system_tag, comm);
count++;
}
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);
- XBT_DEBUG("<%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 system_tag = 889;
int i, rank, size, err, count;
- MPI_Aint lb;
- MPI_Aint sendextent = 0;
- MPI_Aint recvextent = 0;
+ 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);
- err = smpi_datatype_extent(sendtype, &lb, &sendextent);
- err = smpi_datatype_extent(recvtype, &lb, &recvextent);
+ err = smpi_datatype_extent(sendtype, &lb, &sendext);
+ err = smpi_datatype_extent(recvtype, &lb, &recvext);
/* Local copy from self */
err =
- smpi_datatype_copy(&((char *) sendbuf)[senddisps[rank] * sendextent],
+ smpi_datatype_copy((char *)sendbuf + senddisps[rank] * sendext,
sendcounts[rank], sendtype,
- &((char *) recvbuf)[recvdisps[rank] * recvextent],
+ (char *)recvbuf + recvdisps[rank] * recvext,
recvcounts[rank], recvtype);
if (err == MPI_SUCCESS && size > 1) {
/* Initiate all send/recv to/from others. */
continue;
}
requests[count] =
- smpi_irecv_init(&((char *) recvbuf)[recvdisps[i] * recvextent],
+ smpi_irecv_init((char *)recvbuf + recvdisps[i] * recvext,
recvcounts[i], recvtype, i, system_tag, comm);
count++;
}
continue;
}
requests[count] =
- smpi_isend_init(&((char *) sendbuf)[senddisps[i] * sendextent],
+ smpi_isend_init((char *)sendbuf + senddisps[i] * sendext,
sendcounts[i], sendtype, i, system_tag, comm);
count++;
}