void test_2D()
{
int i;
int src, dst;
int ierr;
double *buf;
void *ptr[MAXPROC], *get_ptr[MAXPROC];
/* find who I am and the dst process */
src = me;
#ifdef MALLOC_LOC
if(me == 0) {
buf = (double *)ARMCI_Malloc_local(SIZE * SIZE * sizeof(double));
assert(buf != NULL);
}
#else
if(me == 0) {
buf = (double *)malloc(SIZE * SIZE * sizeof(double));
assert(buf != NULL);
}
#endif
ierr = ARMCI_Malloc(ptr, (SIZE * SIZE * sizeof(double)));
assert(ierr == 0); assert(ptr[me]);
ierr = ARMCI_Malloc(get_ptr, (SIZE * SIZE * sizeof(double)));
assert(ierr == 0); assert(get_ptr[me]);
/* ARMCI - initialize the data window */
fill_array(ptr[me], SIZE*SIZE, me);
fill_array(get_ptr[me], SIZE*SIZE, me);
MP_BARRIER();
/* only the proc 0 doest the work */
/* print the title */
if(me == 0) {
if(!CHECK_RESULT){
printf(" section get put");
printf(" acc\n");
printf("bytes loop sec MB/s sec MB/s");
printf(" sec MB/s\n");
printf("------- ------ -------- -------- -------- --------");
printf(" -------- --------\n");
fflush(stdout);
}
for(i=0; i<CHUNK_NUM; i++) {
int loop;
int bytes = chunk[i] * chunk[i] * sizeof(double);
double t_get = 0, t_put = 0, t_acc = 0;
double latency_get, latency_put, latency_acc;
double bandwidth_get, bandwidth_put, bandwidth_acc;
loop = SIZE / chunk[i];
if(loop<2)loop=2;
for(dst=1; dst<nproc; dst++) {
/* strided get */
fill_array(buf, SIZE*SIZE, me*10);
t_get += time_get((double *)(get_ptr[dst]), (double *)buf,
chunk[i], loop, dst, 1);
/* strided put */
fill_array(buf, SIZE*SIZE, me*10);
t_put += time_put((double *)buf, (double *)(ptr[dst]),
chunk[i], loop, dst, 1);
/* strided acc */
fill_array(buf, SIZE*SIZE, me*10);
t_acc += time_acc((double *)buf, (double *)(ptr[dst]),
chunk[i], loop, dst, 1);
}
latency_get = t_get/(nproc - 1);
latency_put = t_put/(nproc - 1);
latency_acc = t_acc/(nproc - 1);
bandwidth_get = (bytes * (nproc - 1) * 1e-6)/t_get;
bandwidth_put = (bytes * (nproc - 1) * 1e-6)/t_put;
bandwidth_acc = (bytes * (nproc - 1) * 1e-6)/t_acc;
/* print */
if(!CHECK_RESULT)printf("%d\t%d\t%.2e %.2e %.2e %.2e %.2e %.2e\n",
bytes, loop, latency_get, bandwidth_get,
latency_put, bandwidth_put, latency_acc, bandwidth_acc);
}
}
else sleep(3);
ARMCI_AllFence();
MP_BARRIER();
/* cleanup */
ARMCI_Free(get_ptr[me]);
ARMCI_Free(ptr[me]);
#ifdef MALLOC_LOC
if(me == 0) ARMCI_Free_local(buf);
#else
if(me == 0) free(buf);
#endif
}
void test_2D()
{
int i;
int src, dst;
int ierr;
double *buf;
void *ptr[MAXPROC], *get_ptr[MAXPROC];
/* find who I am and the dst process */
src = me;
if(me == 0) {
buf = (double *)malloc(SIZE * SIZE * sizeof(double));
assert(buf != NULL);
}
ierr = A1_Alloc_segment(&ptr[me], (SIZE * SIZE * sizeof(double)));
assert(ierr == 0);
ierr = A1_Exchange_segments(A1_GROUP_WORLD, ptr);
assert(ierr == 0);
ierr = A1_Alloc_segment(&get_ptr[me], (SIZE * SIZE * sizeof(double)));
assert(ierr == 0);
ierr = A1_Exchange_segments(A1_GROUP_WORLD, get_ptr);
assert(ierr == 0);
/* A1 - initialize the data window */
fill_array(ptr[me], SIZE*SIZE, me);
fill_array(get_ptr[me], SIZE*SIZE, me);
A1_Barrier_group(A1_GROUP_WORLD);
/* only the proc 0 doest the work */
/* print the title */
if(me == 0) {
if(!CHECK_RESULT){
printf(" section get put");
printf(" acc\n");
printf("bytes loop usec MB/s usec MB/s");
printf(" usec MB/s\n");
printf("------- ------ -------- -------- -------- --------");
printf(" -------- --------\n");
fflush(stdout);
}
for(i=0; i<CHUNK_NUM; i++) {
int loop;
int bytes = chunk[i] * chunk[i] * sizeof(double);
double t_get = 0, t_put = 0, t_acc = 0;
double latency_get, latency_put, latency_acc;
double bandwidth_get, bandwidth_put, bandwidth_acc;
loop = SIZE / chunk[i];
if(loop<2)loop=2;
for(dst=1; dst<nproc; dst++) {
/* strided get */
fill_array(buf, SIZE*SIZE, me*10);
t_get += time_get((double *)(get_ptr[dst]), (double *)buf,
chunk[i], loop, dst, 1);
/* strided put */
fill_array(buf, SIZE*SIZE, me*10);
t_put += time_put((double *)buf, (double *)(ptr[dst]),
chunk[i], loop, dst, 1);
/* strided acc */
fill_array(buf, SIZE*SIZE, me*10);
t_acc += time_acc((double *)buf, (double *)(ptr[dst]),
chunk[i], loop, dst, 1);
}
latency_get = t_get/(nproc - 1);
latency_put = t_put/(nproc - 1);
latency_acc = t_acc/(nproc - 1);
bandwidth_get = (bytes * (nproc - 1) * 1e-6)/t_get;
bandwidth_put = (bytes * (nproc - 1) * 1e-6)/t_put;
bandwidth_acc = (bytes * (nproc - 1) * 1e-6)/t_acc;
/* print */
if(!CHECK_RESULT)
printf("%d\t%d\t %7.2lf %9.2lf %9.2lf %9.2lf %9.2lf %9.2lf\n",
bytes, loop, latency_get*1000000, bandwidth_get,
latency_put*1000000, bandwidth_put, latency_acc*1000000, bandwidth_acc);
}
}
else sleep(60);
A1_Flush_group(A1_GROUP_WORLD);
A1_Barrier_group(A1_GROUP_WORLD);
/* cleanup */
A1_Release_segments(A1_GROUP_WORLD, get_ptr[me]);
A1_Free_segment(get_ptr[me]);
A1_Release_segments(A1_GROUP_WORLD, ptr[me]);
A1_Free_segment(ptr[me]);
if(me == 0) free(buf);
}
int time_push(lua_State *L, apr_time_t time)
{
time_put(L, time);
return 1;
}
