1
1
openmpi/ompi/mca/coll/libnbc/nbc_iallgatherv.c
Torsten Hoefler 6b22641669 added LibNBC (http://www.unixer.de/NBC) as collv1 (blocking) component.
I know it does not make much sense but one can play around with the
performance. Numbers are available at http://www.unixer.de/research/nbcoll/perf/.
This is the first step towards collv2. Next step includes the addition
of non-blocking functions to the MPI-Layer and the collv1 interface.

It implements all MPI-1 collective algorithms in a non-blocking manner.
However, the collv1 interface does not allow non-blocking collectives so
that all collectives are used blocking by the ompi-glue layer.

I wanted to add LibNBC as a separate subdirectory, but I could not
convince the buildsystem (and had not the time). So the component looks
pretty messy. It would be great if somebody could explain me how to move
all nbc*{c,h}, and {hb,dict}*{c,h} to a seperate subdirectory.

It's .ompi_ignored because I did not test it exhaustively yet.

This commit was SVN r11401.
2006-08-24 16:47:18 +00:00

71 строка
3.0 KiB
C

#include "nbc.h"
/* an allgatherv schedule can not be cached easily because the contents
* ot the recvcounts array may change, so a comparison of the address
* would not be sufficient ... we simply do not cache it */
/* simple linear MPI_Iallgatherv
* the algorithm uses p-1 rounds
* first round:
* each node sends to it's left node (rank+1)%p sendcount elements
* each node begins with it's right node (rank-11)%p and receives from it recvcounts[(rank+1)%p] elements
* second round:
* each node sends to node (rank+2)%p sendcount elements
* each node receives from node (rank-2)%p recvcounts[(rank+2)%p] elements */
int NBC_Iallgatherv(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int *recvcounts, int *displs, MPI_Datatype recvtype, MPI_Comm comm, NBC_Handle *handle) {
int rank, p, res, r, speer, rpeer;
MPI_Aint rcvext, sndext;
NBC_Schedule *schedule;
char *rbuf, *sbuf, inplace;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(handle, comm);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Type_extent(sendtype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
schedule = malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc() (%i)\n", res); return res; }
handle->tmpbuf=NULL;
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create, (%i)\n", res); return res; }
if(!((rank == 0) && inplace)) {
/* copy my data to receive buffer */
rbuf = ((char *)recvbuf) + (displs[rank]*rcvext);
NBC_Copy(sendbuf, sendcount, sendtype, rbuf, recvcounts[rank], recvtype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
sbuf = ((char *)recvbuf) + (displs[rank]*rcvext);
/* do p-1 rounds */
for(r=1;r<p;r++) {
speer = (rank+r)%p;
rpeer = (rank-r+p)%p;
rbuf = ((char *)recvbuf) + (displs[rpeer]*rcvext);
res = NBC_Sched_recv(rbuf, false, recvcounts[rpeer], recvtype, rpeer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
/* send to rank r - not from the sendbuf to optimize MPI_IN_PLACE */
res = NBC_Sched_send(sbuf, false, recvcounts[0], recvtype, speer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}