/* smpi_coll.c -- various optimized routing for collectives */
-/* Copyright (c) 2009, 2010. The SimGrid Team.
+/* Copyright (c) 2009-2014. The SimGrid Team.
* All rights reserved. */
/* This program is free software; you can redistribute it and/or modify it
#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 */
+};
+
+
+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[] = {
- {"ompi",
+ {"default",
"Ompi alltoall default collective",
- smpi_coll_tuned_alltoall_ompi},
-
- {"2dmesh",
- "Alltoall 2dmesh collective",
- smpi_coll_tuned_alltoall_2dmesh},
- {"3dmesh",
- "Alltoall 3dmesh collective",
- smpi_coll_tuned_alltoall_3dmesh},
- /*{"bruck",
- "Alltoall Bruck collective",
- smpi_coll_tuned_alltoall_bruck},*/
- {"pair",
- "Alltoall pair collective",
- smpi_coll_tuned_alltoall_pair},
- {"pair_light_barrier",
- "Alltoall pair_light_barrier collective",
- smpi_coll_tuned_alltoall_pair_light_barrier},
- {"pair_mpi_barrier",
- "Alltoall pair_mpi_barrier collective",
- smpi_coll_tuned_alltoall_pair_mpi_barrier},
- {"rdb",
- "Alltoall rdb collective",
- smpi_coll_tuned_alltoall_rdb},
- {"ring",
- "Alltoall ring collective",
- smpi_coll_tuned_alltoall_ring},
- {"ring_light_barrier",
- "Alltoall ring_light_barrier collective",
- smpi_coll_tuned_alltoall_ring_light_barrier},
- {"ring_light_barrier",
- "Alltoall ring_light_barrier collective",
- smpi_coll_tuned_alltoall_ring_light_barrier},
- {"ring_mpi_barrier",
- "Alltoall ring_mpi_barrier collective",
- smpi_coll_tuned_alltoall_ring_mpi_barrier},
- {"ring_one_barrier",
- "Alltoall ring_one_barrier collective",
- smpi_coll_tuned_alltoall_ring_one_barrier},
- {"simple",
- "Alltoall simple collective",
- smpi_coll_tuned_alltoall_simple},
-
+ 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},
- {"pairwise",
- "Alltoall pairwise (SG) collective",
- smpi_coll_tuned_alltoall_pairwise},
+ {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_allgather_description[] = {
+s_mpi_coll_description_t mpi_coll_bcast_description[] = {
{"default",
- "allgather default collective",
- smpi_mpi_gather},
+ "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 find_coll_description(s_mpi_coll_description_t * table,
- const char *name)
+ 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;
}
- name_list = strdup(table[0].name);
+
+ 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,
strcat(name_list, ", ");
strcat(name_list, table[i].name);
}
- xbt_die("Model '%s' is invalid! Valid models are: %s.", name, name_list);
+ 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;
static void tree_bcast(void *buf, int count, MPI_Datatype datatype,
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;
}
smpi_mpi_startall(tree->numChildren, requests);
smpi_mpi_waitall(tree->numChildren, requests, MPI_STATUS_IGNORE);
+ for(i = 0; i < tree->numChildren; i++) {
+ if(requests[i]!=MPI_REQUEST_NULL) smpi_mpi_request_free(&requests[i]);
+ }
xbt_free(requests);
}
static void tree_antibcast(void *buf, int count, MPI_Datatype datatype,
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;
}
smpi_mpi_startall(tree->numChildren, requests);
smpi_mpi_waitall(tree->numChildren, requests, MPI_STATUS_IGNORE);
+ for(i = 0; i < tree->numChildren; i++) {
+ if(requests[i]!=MPI_REQUEST_NULL) smpi_mpi_request_free(&requests[i]);
+ }
xbt_free(requests);
}
free_tree(tree);
}
-int smpi_coll_tuned_alltoall_ompi(void *sendbuf, int sendcount,
+int smpi_coll_tuned_alltoall_ompi2(void *sendbuf, int sendcount,
MPI_Datatype sendtype, void *recvbuf,
int recvcount, MPI_Datatype recvtype,
MPI_Comm comm)
recvcount, recvtype, comm);
} else {
return
- smpi_coll_tuned_alltoall_pairwise(sendbuf, sendcount, sendtype,
- recvbuf, recvcount, recvtype,
- comm);
+ smpi_coll_tuned_alltoall_ring(sendbuf, sendcount, sendtype,
+ recvbuf, recvcount, recvtype,
+ comm);
}
}
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
XBT_DEBUG("<%d> algorithm alltoall_bruck() called.", rank);
- err = smpi_datatype_extent(sendtype, &lb, &sendext);
- err = smpi_datatype_extent(recvtype, &lb, &recvext);
+ 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,
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;
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, &sendext);
- err = smpi_datatype_extent(recvtype, &lb, &recvext);
+ 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,
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;
}
-/**
- * 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,
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
XBT_DEBUG("<%d> algorithm basic_alltoallv() called.", rank);
- err = smpi_datatype_extent(sendtype, &lb, &sendext);
- err = smpi_datatype_extent(recvtype, &lb, &recvext);
+ 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,
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;