|
<![CDATA[/*\
* pingpong.c -- hcc pingpong.c -o pingpong [-lmpi]
*
* this program does just as the name suggests -- pingpongs a message
* between two MPI nodes.
*
* Written by Pete Rijks <prijks@nd.edu> for LAM performance testing
*
\*/
#include <stdlib.h>
#include <stdio.h>
#include "mpi.h"
#include <sys/time.h>
/* max message size to test */
#define MAXCOUNT (1048576 * 8)
/* number of tests per message size */
#define NUMTESTS 50
/* number of initial test per message size to discard results from */
#define IGNORE 10
typedef unsigned long long int pptimer_t;
#ifdef __GNUC__
__inline__
#endif
pptimer_t rdtsc();
int
main(int argc, char **argv)
{
int i, j, rank, size;
int to, from, tag;
int count;
pptimer_t total, timer_begin, timer_end, timer_diff, avg;
pptimer_t *avgs;
char *message;
MPI_Status status;
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
avgs = (pptimer_t*) malloc (24 * sizeof(pptimer_t));
message = (char*) malloc (MAXCOUNT * sizeof(char));
from = 1 - rank;
to = 1 - rank;
tag = 666;
if (rank == 0) {
for (count = 1, j = 0; count <= MAXCOUNT; count*=2, j++) {
fprintf(stderr,"%10d",count);
MPI_Barrier(MPI_COMM_WORLD);
total = 0;
for (i = 0; i < NUMTESTS; i++) {
fprintf(stderr,".");
timer_begin = rdtsc();
MPI_Send(message, count, MPI_CHAR, to, tag, MPI_COMM_WORLD);
MPI_Recv(message, count, MPI_CHAR, from, tag, MPI_COMM_WORLD, &status);
timer_end = rdtsc();
timer_diff = timer_end - timer_begin;
if (i > IGNORE)
total += timer_diff;
}
fprintf(stderr,"\n");
avg = total / (NUMTESTS - IGNORE);
avgs[j] = avg;
}
/* output is matlab-friendly */
printf("x = [ ");
for (count = 1; count <= MAXCOUNT; count*=2)
printf("%d ",count);
printf("]; \ncycles = [ ");
for (i = 0; i < j; i++) {
/* grr... why can't long long ints print out the same everywhere? */
#ifdef __GNUC__
printf("%Ld ",avgs[i]);
#else
printf("%lld ",avgs[i]);
#endif
}
printf("];\n");
printf("loglog(x,cycles,'x-');\n");
} else if (rank == 1) {
for (count = 1; count <= MAXCOUNT; count*=2) {
MPI_Barrier(MPI_COMM_WORLD);
for (i = 0; i < NUMTESTS; i++) {
MPI_Recv(message, count, MPI_CHAR, from, tag, MPI_COMM_WORLD, &status);
MPI_Send(message, count, MPI_CHAR, to, tag, MPI_COMM_WORLD);
}
}
} else {
/* no point in running this on more than two nodes... */
fprintf(stderr,"notice: pingpong requires only two nodes to run on.\n");
/* but if somebody does run it on more than two -- it'll hang *
* if the other nodes don't participate in MPI_Barrier's... */
for (count = 1; count <= MAXCOUNT; count*=2)
MPI_Barrier(MPI_COMM_WORLD);
}
/* clean up */
fflush(stdout);
free(avgs);
free(message);
MPI_Finalize();
return 0;
}
/* hi-res timing -- the asm below returns the number of instructions executed
* since system boot up. Pretty hi-res if you ask me. x86-only, tho.
* on suns we use gethrtime (get high resolution time) instead...
*/
#ifdef __GNUC__
__inline__
#endif
pptimer_t rdtsc()
{
pptimer_t x;
#ifdef __GNUC__
__asm__ volatile (".byte 0x0f, 0x31" : "=A" (x));
#else
x = gethrtime();
#endif
return x;
}
]]>
|
|
