MPICH on Linux with PGI can be very tricky to set up properly. There are a few things you
may want to try.
1) Do the simple control experiment of testing a basic program. AMBER is a demanding
MPI application, and can break MPICH in some interesting ways. But if
you can't even run a simple MPI application, then you know to look elsewhere.
2) Try it with AMBER and MPICH compiled with gcc/g77 (make sure both are compiled using the same compiler)
3) If you're using the shared memory communicator (IE on a dual processor)
try setting the following in .bashrc:
export P4_GLOBALMEMSIZE=1000000
I use the following app to test my cluster. It basically just saturates the network
by doing send/recv between pairs of machines. Run it with an even number of processors, 2 or more.
#include <math.h>
#include <malloc.h>
#include <stdio.h>
#include "mpi.h"
#define SIZE 4096
int main(int argc,char *argv[]) {
MPI_Status status;
int me, np;
int i, j;
double t0, t1, overhead, dt;
char *data;
char hostname[256];
int num_cycles;
MPI_Init( &argc, &argv );
MPI_Comm_rank( MPI_COMM_WORLD, &me );
MPI_Comm_size( MPI_COMM_WORLD, &np );
if (np % 2 != 0) {
if (me == 0) {
printf("Error: # cpus must be multiple of 2!\n");
}
exit(1);
}
if (me % 2 == 0) {
t0 = MPI_Wtime();
num_cycles = 0;
}
data = (char *)calloc(sizeof(char) * SIZE, 0);
if (me==0) printf("Starting\n");
while(1) {
if (me % 2 == 0) {
MPI_Send(data,SIZE,MPI_CHAR,me+1,1,MPI_COMM_WORLD);
}
else {
MPI_Recv(data,SIZE,MPI_CHAR,me-1,1,MPI_COMM_WORLD, &status);
}
if (me % 2 == 0) {
t1 = MPI_Wtime();
num_cycles++;
if (t1-t0 > 5.0) {
printf("Node %d to %d: throughput %5.2f Mbytes/sec\n", me,me+1,(num_cycles*SIZE)/(t1-t0)/1.e6);
fflush(stdout);
num_cycles = 0;
t0 = MPI_Wtime();
}
}
}
MPI_Finalize();
}
Received on Fri May 26 2000 - 08:02:18 PDT
