int main(int argc, char **argv)
{
int k,i;
double **myptrs[10];
double t0,t1,tget=0,tnbget=0,tput=0,tnbput=0,tnbwait=0,t2=0;
#if PORTALS
ARMCI_NetInit();
#endif
MPI_Init(&argc,&argv);
MPI_Comm_rank(MPI_COMM_WORLD,&me);
MPI_Comm_size(MPI_COMM_WORLD,&nprocs);
ARMCI_Init();
ARMCI_Init();
for(k=0;k<10;k++){
myptrs[k] = (double **)malloc(sizeof(double *)*nprocs);
ARMCI_Malloc((void **)myptrs[k],400000*LOOP*sizeof(double));
for(i=0;i<LOOP;i++)myptrs[k][me][i]=me+0.414;
MPI_Barrier(MPI_COMM_WORLD);
for(i=0;i<LOOP;i++){
ARMCI_Get(myptrs[k][(me+1)%nprocs]+i,myptrs[k][me]+i,sizeof(double),(me+1)%nprocs);
/*if(myptrs[k][me][i]!=0.414+(me+1)%nprocs)ARMCI_Error("errr",myptrs[k][me][i]);*/
}
t0=t1=tget=tnbget=tput=tnbput=tnbwait=t2=0;
t0 = MPI_Wtime();
for(i=0;i<LOOP;i++){
ARMCI_Get(myptrs[k][(me+1)%nprocs]+i,myptrs[k][me]+i,sizeof(double),(me+1)%nprocs);
}
t1 = MPI_Wtime();
printf("\nGet Latency=%lf\n",1e6*(t1-t0)/LOOP);fflush(stdout);
t1=t0=0;
for(i=0;i<LOOP;i++){
armci_hdl_t nbh;
ARMCI_INIT_HANDLE(&nbh);
t0 = MPI_Wtime();
ARMCI_NbGet(myptrs[k][(me+1)%nprocs]+i,myptrs[k][me]+i,sizeof(double),(me+1)%nprocs,&nbh);
t1 = MPI_Wtime();
ARMCI_Wait(&nbh);
t2 = MPI_Wtime();
tnbget+=(t1-t0);
tnbwait+=(t2-t1);
}
printf("\nNb Get Latency=%lf Nb Wait=%lf\n",1e6*tnbget/LOOP,1e6*tnbwait/LOOP);fflush(stdout);
MPI_Barrier(MPI_COMM_WORLD);
}
for(k=0;k<10;k++)ARMCI_Free(myptrs[k][me]);
MPI_Barrier(MPI_COMM_WORLD);
ARMCI_Finalize();
ARMCI_Finalize();
MPI_Finalize();
}
void test_aggregate(int dryrun) {
int i, j, rc, bytes, elems[2] = {MAXPROC, MAXELEMS};
double *ddst_put[MAXPROC];
double *ddst_get[MAXPROC];
double *dsrc[MAXPROC];
armci_hdl_t aggr_hdl_put[MAXPROC];
armci_hdl_t aggr_hdl_get[MAXPROC];
armci_hdl_t hdl_put[MAXELEMS];
armci_hdl_t hdl_get[MAXELEMS];
armci_giov_t darr;
void *src_ptr[MAX_REQUESTS], *dst_ptr[MAX_REQUESTS];
int start = 0, end = 0;
double start_time;
create_array((void**)ddst_put, sizeof(double),2, elems);
create_array((void**)ddst_get, sizeof(double),2, elems);
create_array((void**)dsrc, sizeof(double),1, &elems[1]);
for(i=0; i<elems[1]; i++) dsrc[me][i]=i*1.001*(me+1);
for(i=0; i<elems[0]*elems[1]; i++) {
ddst_put[me][i]=0.0;
ddst_get[me][i]=0.0;
}
MP_BARRIER();
/* only proc 0 does the work */
if(me == 0) {
if(!dryrun)printf("Transferring %d doubles (Not an array of %d doubles)\n", MAXELEMS, MAXELEMS);
/* initializing non-blocking handles */
for(i=0; i<elems[1]; i++) ARMCI_INIT_HANDLE(&hdl_put[i]);
for(i=0; i<elems[1]; i++) ARMCI_INIT_HANDLE(&hdl_get[i]);
/* aggregate handles */
for(i=0; i<nproc; i++) ARMCI_INIT_HANDLE(&aggr_hdl_put[i]);
for(i=0; i<nproc; i++) ARMCI_INIT_HANDLE(&aggr_hdl_get[i]);
for(i=0; i<nproc; i++) ARMCI_SET_AGGREGATE_HANDLE(&aggr_hdl_put[i]);
for(i=0; i<nproc; i++) ARMCI_SET_AGGREGATE_HANDLE(&aggr_hdl_get[i]);
bytes = sizeof(double);
/* **************** PUT **************** */
/* register put */
start_time=MP_TIMER();
start = 0; end = elems[1];
for(i=1; i<nproc; i++) {
for(j=start; j<end; j++) {
ARMCI_NbPutValueDouble(dsrc[me][j], &ddst_put[i][me*elems[1]+j], i,
&hdl_put[j]);
}
for(j=start; j<end; j++) ARMCI_Wait(&hdl_put[j]);
}
if(!dryrun)printf("%d: Value Put time = %.2es\n", me, MP_TIMER()-start_time);
/* vector put */
start_time=MP_TIMER();
for(i=1; i<nproc; i++) {
for(j=start; j<end; j++) {
src_ptr[j] = (void *)&dsrc[me][j];
dst_ptr[j] = (void *)&ddst_put[i][me*elems[1]+j];
}
darr.src_ptr_array = src_ptr;
darr.dst_ptr_array = dst_ptr;
darr.bytes = sizeof(double);
darr.ptr_array_len = elems[1];
if((rc=ARMCI_NbPutV(&darr, 1, i, &hdl_put[i])))
ARMCI_Error("armci_nbputv failed\n",rc);
}
for(i=1; i<nproc; i++) ARMCI_Wait(&hdl_put[i]);
if(!dryrun)printf("%d: Vector Put time = %.2es\n", me, MP_TIMER()-start_time);
/* regular put */
start_time=MP_TIMER();
for(i=1; i<nproc; i++) {
for(j=start; j<end; j++) {
if((rc=ARMCI_NbPut(&dsrc[me][j], &ddst_put[i][me*elems[1]+j], bytes,
i, &hdl_put[j])))
ARMCI_Error("armci_nbput failed\n",rc);
}
for(j=start; j<end; j++) ARMCI_Wait(&hdl_put[j]);
}
if(!dryrun)printf("%d: Regular Put time = %.2es\n", me, MP_TIMER()-start_time);
/* aggregate put */
start_time=MP_TIMER();
for(i=1; i<nproc; i++) {
for(j=start; j<end; j++) {
if((rc=ARMCI_NbPut(&dsrc[me][j], &ddst_put[i][me*elems[1]+j], bytes,
i, &aggr_hdl_put[i])))
ARMCI_Error("armci_nbput failed\n",rc);
}
}
for(i=1; i<nproc; i++) ARMCI_Wait(&aggr_hdl_put[i]);
if(!dryrun)printf("%d: Aggregate Put time = %.2es\n\n", me, MP_TIMER()-start_time);
/* **************** GET **************** */
/* vector get */
start_time=MP_TIMER();
for(i=1; i<nproc; i++) {
for(j=start; j<end; j++) {
src_ptr[j] = (void *)&dsrc[i][j];
dst_ptr[j] = (void *)&ddst_get[me][i*elems[1]+j];
}
darr.src_ptr_array = src_ptr;
darr.dst_ptr_array = dst_ptr;
darr.bytes = sizeof(double);
darr.ptr_array_len = elems[1];
if((rc=ARMCI_NbGetV(&darr, 1, i, &hdl_get[i])))
ARMCI_Error("armci_nbgetv failed\n",rc);
ARMCI_Wait(&hdl_get[i]);
}
if(!dryrun)printf("%d: Vector Get time = %.2es\n", me, MP_TIMER()-start_time);
/* regular get */
start_time=MP_TIMER();
for(i=1; i<nproc; i++) {
for(j=start; j<end; j++) {
if((rc=ARMCI_NbGet(&dsrc[i][j], &ddst_get[me][i*elems[1]+j], bytes,
i, &hdl_get[j])))
ARMCI_Error("armci_nbget failed\n",rc);
}
for(j=start; j<end; j++) ARMCI_Wait(&hdl_get[j]);
}
if(!dryrun)printf("%d: Regular Get time = %.2es\n", me, MP_TIMER()-start_time);
/* aggregate get */
start_time=MP_TIMER();
for(i=1; i<nproc; i++) {
for(j=start; j<end; j++) {
ARMCI_NbGet(&dsrc[i][j], &ddst_get[me][i*elems[1]+j], bytes,
i, &aggr_hdl_get[i]);
}
}
for(i=1; i<nproc; i++) ARMCI_Wait(&aggr_hdl_get[i]);
if(!dryrun)printf("%d: Aggregate Get time = %.2es\n", me, MP_TIMER()-start_time);
}
MP_BARRIER();
ARMCI_AllFence();
MP_BARRIER();
/* Verify */
if(!(me==0))
for(j=0; j<elems[1]; j++) {
if( ARMCI_ABS(ddst_put[me][j]-j*1.001) > 0.1) {
ARMCI_Error("aggregate put failed...1", 0);
}
}
MP_BARRIER();
if(!dryrun)if(me==0) printf("\n aggregate put ..O.K.\n"); fflush(stdout);
if(me==0) {
for(i=1; i<nproc; i++) {
for(j=0; j<elems[1]; j++) {
if( ARMCI_ABS(ddst_get[me][i*elems[1]+j]-j*1.001*(i+1)) > 0.1) {
ARMCI_Error("aggregate get failed...1", 0);
}
}
}
}
MP_BARRIER();
if(!dryrun)if(me==0) printf(" aggregate get ..O.K.\n"); fflush(stdout);
ARMCI_AllFence();
MP_BARRIER();
if(!dryrun)if(me==0){printf("O.K.\n"); fflush(stdout);}
destroy_array((void **)ddst_put);
destroy_array((void **)ddst_get);
destroy_array((void **)dsrc);
}
int main(int argc, char **argv) {
int i;
double **myptrs;
double t0, t1, tnbget=0, tnbwait=0, t2=0;
MP_INIT(argc,argv);
ARMCI_Init();
MP_PROCS(&nprocs);
MP_MYID(&me);
if (nprocs < 2)
ARMCI_Error("This program requires at least to processes", 1);
myptrs = (double **)malloc(sizeof(double *)*nprocs);
ARMCI_Malloc((void **)myptrs, LOOP*sizeof(double));
MP_BARRIER();
if(me == 0) {
for(i = 0; i < 10; i++) {
// This is a bug:
// ARMCI_Get(myptrs[me]+i,myptrs[me+1]+i,sizeof(double),me+1);
ARMCI_Get(myptrs[me+1]+i, myptrs[me]+i, sizeof(double), me+1);
}
t0 = MP_TIMER();
for(i = 0; i < LOOP; i++) {
// This is a bug:
// ARMCI_Get(myptrs[me]+i,myptrs[me+1]+i,sizeof(double),me+1);
ARMCI_Get(myptrs[me+1]+1, myptrs[me]+i, sizeof(double), me+1);
}
t1 = MP_TIMER();
printf("\nGet Latency=%lf\n", 1e6*(t1-t0)/LOOP);
fflush(stdout);
t1 = t0 = 0;
for(i = 0; i < LOOP; i++) {
armci_hdl_t nbh;
ARMCI_INIT_HANDLE(&nbh);
t0 = MP_TIMER();
//ARMCI_NbGet(myptrs[me]+i, myptrs[me+1]+i, sizeof(double), me+1, &nbh);
ARMCI_NbGet(myptrs[me+1]+i, myptrs[me]+i, sizeof(double), me+1, &nbh);
t1 = MP_TIMER();
ARMCI_Wait(&nbh);
t2 = MP_TIMER();
tnbget += (t1-t0);
tnbwait += (t2-t1);
}
printf("\nNb Get Latency=%lf Nb Wait=%lf\n",1e6*tnbget/LOOP,1e6*tnbwait/LOOP);fflush(stdout);
}
else
sleep(1);
MP_BARRIER();
ARMCI_Finalize();
MP_FINALIZE();
return 0;
}
static void sparse_multiply(int n, int non_zero, int *row_ind, int **col_ind,
double **values, double **vec, double **svec) {
int i, j, k, num_elements, offset, *tmp_indices;
double start_time, comm_time, comp_time, v, vec_local[COL];
int start, end, prev, nrows, idx;
#if 0
/* ---- Sequential Case ---- */
for(i=0; i<n; i++) {
svec[me][i] = 0;
for(k=row_ind[i]; k<row_ind[i+1]; k++) {
j = col_ind[me][k];
v = values[me][k];
svec[me][i] += v*vec[me][j];
printf("%.1f %.1f\n", v, vec[me][j]);
}
}
for(i=0; i<n; i++) printf("%.1f ", svec[me][i]);
printf("\n");
#else
num_elements = proc_nz_list[me];
printf("num_elements = %d\n", num_elements);
tmp_indices = (int *)malloc(num_elements*sizeof(int));
for(i=0; i<num_elements; i++) tmp_indices[i] = col_ind[me][i];
qsort(tmp_indices, num_elements, sizeof(int), compare);
start_time = armci_timer();
/* get the required portion of vector you need to local array */
start = prev = tmp_indices[0];
for(i=1; i<num_elements; i++) {
if(tmp_indices[i]>prev+1) {
end = prev;
get_data(n, start, end, vec_local, vec);
start = prev = tmp_indices[i];
}
else prev = tmp_indices[i];
}
get_data(n, start, prev, vec_local, vec);
#if 1
if(count>=0) for(i=0; i<=count; i++) ARMCI_Wait(&gHandle[i]);
#endif
comm_time = armci_timer() - start_time;
start_time = armci_timer();
/* Perform Matrix-Vector multiply and store the result in
solution vector - "svec[]" */
if(me==0) {
nrows = proc_row_list[me];
offset = row_ind[0];
}
else {
nrows = proc_row_list[me]-proc_row_list[me-1];
offset = row_ind[proc_row_list[me-1]];
}
/* printf("%d: My total Work = %d\n", me, nrows); */
for(i=0; i<nrows; i++) { /* loop over rows owned by me */
svec[me][i] = 0;
if(me==0) idx = i;
else idx = proc_row_list[me-1] + i;
for(k=row_ind[idx]; k<row_ind[idx+1]; k++) {
j = col_ind[me][k-offset];
v = values[me][k-offset];
svec[me][i] += v*vec_local[j];
}
}
comp_time = armci_timer()-start_time;
printf("%d: %f + %f = %f (count = %d)\n", me, comm_time, comp_time,
comm_time+comp_time, count+1);
#endif
}
int main(int argc, char *argv[])
{
size_t i, rank, nranks, msgsize, dest;
size_t iterations, max_msgsize;
int bufsize;
double **buffer;
double t_start, t_stop, t_total, d_total;
double expected, bandwidth;
int provided;
armci_hdl_t handle;
max_msgsize = MAX_MSGSIZE;
MPI_Init_thread(&argc, &argv, MPI_THREAD_MULTIPLE, &provided);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
ARMCI_Init_args(&argc, &argv);
bufsize = max_msgsize * ITERATIONS_LARGE;
buffer = (double **) malloc(sizeof(double *) * nranks);
ARMCI_Malloc((void **) buffer, bufsize);
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
ARMCI_INIT_HANDLE(&handle);
ARMCI_SET_AGGREGATE_HANDLE(&handle);
ARMCI_Barrier();
if (rank == 0)
{
printf("ARMCI_Put Bandwidth in MBPS \n");
printf("%20s %22s \n", "Message Size", "Bandwidth");
fflush(stdout);
dest = 1;
expected = 1 + dest;
for (msgsize = sizeof(double); msgsize <= max_msgsize; msgsize *= 2)
{
if (msgsize <= 16 * 1024) iterations = ITERATIONS_VERYSMALL;
else if (msgsize <= 64 * 1024) iterations = ITERATIONS_SMALL;
else if (msgsize <= 512 * 1024) iterations = ITERATIONS_MEDIUM;
else iterations = ITERATIONS_LARGE;
t_start = MPI_Wtime();
for (i = 0; i < iterations; i++)
{
ARMCI_NbPut((void *) ((size_t) buffer[dest] + (size_t)(i
* msgsize)), (void *) ((size_t) buffer[rank]
+ (size_t)(i * msgsize)), msgsize, dest, &handle);
}
ARMCI_Wait(&handle);
t_stop = MPI_Wtime();
d_total = (iterations * msgsize) / (1024 * 1024);
t_total = t_stop - t_start;
bandwidth = d_total / t_total;
printf("%20d %20.4lf \n", msgsize, bandwidth);
fflush(stdout);
ARMCI_Fence(dest);
}
}
ARMCI_Barrier();
ARMCI_UNSET_AGGREGATE_HANDLE(&handle);
ARMCI_Free((void *) buffer[rank]);
ARMCI_Finalize();
MPI_Finalize();
return 0;
}
void armci_agg_complete(armci_ihdl_t nb_handle, int condition) {
int i, index=0, rc;
/* get the buffer index for this handle */
for(i=ulist.size-1; i>=0; i--) {
index = ulist.index[i];
if(aggr[index]->tag == nb_handle->tag &&
aggr[index]->proc == nb_handle->proc)
break;
}
if(i<0) return; /* implies this handle has no requests at all */
#if 0
printf("%d: Aggregation Complete to remote process %d (%d:%d requests)\n",
armci_me, nb_handle->proc, index, aggr[index]->request_len);
#endif
/* complete the data transfer. NOTE: in LAPI, Non-blocking calls
(followed by wait) performs better than blocking put/get */
if(aggr[index]->request_len) {
switch(nb_handle->op) {
#ifdef LAPI
armci_hdl_t usr_hdl;
case PUT:
ARMCI_INIT_HANDLE(&usr_hdl);
if((rc=ARMCI_NbPutV(aggr[index]->darr, aggr[index]->request_len,
nb_handle->proc, (armci_hdl_t*)&usr_hdl)))
ARMCI_Error("armci_agg_complete: nbputv failed",rc);
ARMCI_Wait((armci_hdl_t*)&usr_hdl);
break;
case GET:
ARMCI_INIT_HANDLE(&usr_hdl);
if((rc=ARMCI_NbGetV(aggr[index]->darr, aggr[index]->request_len,
nb_handle->proc, (armci_hdl_t*)&usr_hdl)))
ARMCI_Error("armci_agg_complete: nbgetv failed",rc);
ARMCI_Wait((armci_hdl_t*)&usr_hdl);
break;
#else
case PUT:
if((rc=ARMCI_PutV(aggr[index]->darr, aggr[index]->request_len,
nb_handle->proc)))
ARMCI_Error("armci_agg_complete: putv failed",rc);
break;
case GET:
if((rc=ARMCI_GetV(aggr[index]->darr, aggr[index]->request_len,
nb_handle->proc)))
ARMCI_Error("armci_agg_complete: getv failed",rc);
break;
#endif
}
}
/* setting request length to zero, as the requests are completed */
aggr[index]->request_len = 0;
aggr[index]->ptr_array_len = 0;
aggr[index]->buf_pos_end = _MAX_AGG_BUFSIZE;
/* If armci_agg_complete() is called ARMCI_Wait(), then unset nb_handle*/
if(condition==UNSET) {
nb_handle->proc = -1;
_armci_agg_update_lists(index);
}
}
int main(int argc, char *argv[])
{
int rank, nranks;
size_t i, msgsize, dest;
size_t iterations, max_msgsize;
int bufsize;
double **buffer;
double t_start, t_stop, t_total, d_total;
double expected, bandwidth;
int provided;
armci_hdl_t handle;
MPI_Init_thread(&argc, &argv, MPI_THREAD_MULTIPLE, &provided);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
max_msgsize = MAX_MSGSIZE;
ARMCI_Init_args(&argc, &argv);
bufsize = max_msgsize * ITERATIONS;
buffer = (double **) malloc(sizeof(double *) * nranks);
ARMCI_Malloc((void **) buffer, bufsize);
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
ARMCI_INIT_HANDLE(&handle);
ARMCI_SET_AGGREGATE_HANDLE(&handle);
ARMCI_Barrier();
if (rank == 0)
{
printf("ARMCI_Get Bandwidth in MBPS \n");
printf("%20s %22s \n", "Message Size", "Bandwidth");
fflush(stdout);
dest = 1;
expected = 1 + dest;
for (msgsize = sizeof(double); msgsize <= max_msgsize; msgsize *= 2)
{
iterations = bufsize/msgsize;
t_start = MPI_Wtime();
for (i = 0; i < iterations; i++)
{
ARMCI_NbGet((void *) ((size_t) buffer[dest] + (size_t)(i
* msgsize)), (void *) ((size_t) buffer[rank]
+ (size_t)(i * msgsize)), msgsize, dest, &handle);
}
ARMCI_Wait(&handle);
t_stop = MPI_Wtime();
d_total = (iterations * msgsize) / (1024 * 1024);
t_total = t_stop - t_start;
bandwidth = d_total / t_total;
printf("%20d %20.4lf \n", msgsize, bandwidth);
fflush(stdout);
#ifdef DATA_VALIDATION
{
for(j=0; j<((iterations*msgsize)/sizeof(double)); j++)
{
if(*(buffer[rank] + j) != expected)
{
printf("Data validation failed At displacement : %d Expected : %lf Actual : %lf \n",
j, expected, *(buffer[rank] + j));
fflush(stdout);
return -1;
}
}
for(j=0; j<bufsize/sizeof(double); j++)
{
*(buffer[rank] + j) = 1.0 + rank;
}
}
#endif
}
}
ARMCI_Barrier();
ARMCI_UNSET_AGGREGATE_HANDLE(&handle);
ARMCI_Free((void *) buffer[rank]);
ARMCI_Finalize();
MPI_Finalize();
return 0;
}
void ARMCI_Gpc_wait(gpc_hdl_t *nbh) {
if(SAMECLUSNODE(nbh->proc))
return;
ARMCI_Wait(&nbh->ahdl);
}
int main(int argc, char *argv[])
{
int i, j, rank, nranks, msgsize, dest;
int dim, iterations;
long bufsize;
double **buffer;
double t_start, t_stop, t_total, d_total, bw;
int count[2], src_stride, trg_stride, stride_level;
int provided;
armci_hdl_t handle;
MPI_Init_thread(&argc, &argv, MPI_THREAD_MULTIPLE, &provided);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
ARMCI_Init_args(&argc, &argv);
bufsize = MAX_DIM * MAX_DIM * sizeof(double);
buffer = (double **) malloc(sizeof(double *) * nranks);
ARMCI_Malloc((void **) buffer, bufsize);
for (i = 0; i < bufsize / sizeof(double); i++)
{
*(buffer[rank] + i) = 1.0 + rank;
}
ARMCI_INIT_HANDLE(&handle);
ARMCI_SET_AGGREGATE_HANDLE(&handle);
ARMCI_Barrier();
if (rank == 0)
{
printf("ARMCI_PutS Bandwidth in MBPS \n");
printf("%30s %22s \n", "Dimensions(array of doubles)", "Latency");
fflush(stdout);
dest = 1;
src_stride = MAX_DIM * sizeof(double);
trg_stride = MAX_DIM * sizeof(double);
stride_level = 1;
for (dim = 1; dim <= MAX_DIM; dim *= 2)
{
count[0] = dim*sizeof(double);
count[1] = dim;
iterations = 10*(MAX_DIM * MAX_DIM)/(dim * dim);
t_start = MPI_Wtime();
for (i = 0; i < iterations; i++)
{
ARMCI_NbPutS((void *) buffer[rank],
&src_stride,
(void *) buffer[dest],
&trg_stride,
count,
stride_level,
dest,
&handle);
}
ARMCI_Wait(&handle);
t_stop = MPI_Wtime();
ARMCI_Fence(1);
char temp[10];
sprintf(temp, "%dX%d", dim, dim);
t_total = t_stop - t_start;
d_total = (dim*dim*sizeof(double)*iterations)/(1024*1024);
bw = d_total/t_total;
printf("%30s %20.2f \n", temp, bw);
fflush(stdout);
}
}
ARMCI_Barrier();
ARMCI_UNSET_AGGREGATE_HANDLE(&handle);
ARMCI_Free((void *) buffer[rank]);
ARMCI_Finalize();
MPI_Finalize();
return 0;
}
void test_perf_nb(int dry_run) {
int i, j, loop, rc, bytes, elems[2] = {MAXPROC, MAXELEMS};
int stride, k=0, ntimes;
double stime, t1, t2, t3, t4, t5, t6, t7, t8, t9;
double *dsrc[MAXPROC], scale=1.0;
armci_hdl_t hdl_get, hdl_put, hdl_acc;
create_array((void**)ddst, sizeof(double),2, elems);
create_array((void**)dsrc, sizeof(double),1, &elems[1]);
if(!dry_run)if(me == 0) {
printf("\n\t\t\tRemote 1-D Array Section\n");
printf("section get nbget wait put nbput ");
printf(" wait acc nbacc wait\n");
printf("------- -------- -------- -------- -------- --------");
printf(" -------- -------- -------- --------\n");
fflush(stdout);
}
for(loop=1; loop<=MAXELEMS; loop*=2, k++) {
elems[1] = loop;
ntimes = (int)sqrt((double)(MAXELEMS/elems[1]));
if(ntimes <1) ntimes=1;
/* -------------------------- SETUP --------------------------- */
/*initializing non-blocking handles,time,src & dst buffers*/
ARMCI_INIT_HANDLE(&hdl_put);
ARMCI_INIT_HANDLE(&hdl_get);
ARMCI_INIT_HANDLE(&hdl_acc);
t1 = t2 = t3 = t4 = t5 = t6 = t7 = t8 = t9 = 0.0;
for(i=0; i<elems[1]; i++) dsrc[me][i]=i*1.001*(me+1);
for(i=0; i<elems[0]*elems[1]; i++) ddst[me][i]=0.0;
MP_BARRIER();
/* bytes transfered */
bytes = sizeof(double)*elems[1];
MP_BARRIER();
/* -------------------------- PUT/GET -------------------------- */
if(me == 0) {
for(i=1; i<nproc; i++) {
stime=MP_TIMER();
for(j=0; j<ntimes; j++)
if((rc=ARMCI_Put(&dsrc[me][0], &ddst[i][me*elems[1]], bytes,i)))
ARMCI_Error("armci_nbput failed\n",rc);
t1 += MP_TIMER()-stime;
}
}
MP_BARRIER(); ARMCI_AllFence(); MP_BARRIER();
if(VERIFY) verify_results(PUT, elems);
for(i=0; i<elems[0]*elems[1]; i++) ddst[me][i]=0.0;
MP_BARRIER();
if(me == 0) {
for(i=1; i<nproc; i++) {
stime=MP_TIMER();
for(j=0; j<ntimes; j++)
if((rc=ARMCI_Get(&dsrc[i][0], &ddst[me][i*elems[1]], bytes,i)))
ARMCI_Error("armci_nbget failed\n",rc);
t4 += MP_TIMER()-stime;
}
}
MP_BARRIER(); ARMCI_AllFence(); MP_BARRIER();
if(VERIFY) verify_results(GET, elems);
for(i=0; i<elems[0]*elems[1]; i++) ddst[me][i]=0.0;
MP_BARRIER();
/* ------------------------ nb PUT/GET ------------------------- */
if(me == 0) {
for(i=1; i<nproc; i++) {
for(j=0; j<ntimes; j++) {
stime=MP_TIMER();
if((rc=ARMCI_NbPut(&dsrc[me][0], &ddst[i][me*elems[1]], bytes,
i, &hdl_put)))
ARMCI_Error("armci_nbput failed\n",rc);
t2 += MP_TIMER()-stime; stime=MP_TIMER();
ARMCI_Wait(&hdl_put);
t3 += MP_TIMER()-stime;
}
}
}
MP_BARRIER(); ARMCI_AllFence(); MP_BARRIER();
if(VERIFY) verify_results(PUT, elems);
for(i=0; i<elems[0]*elems[1]; i++) ddst[me][i]=0.0;
MP_BARRIER();
if(me == 0) {
for(i=1; i<nproc; i++) {
for(j=0; j<ntimes; j++) {
stime=MP_TIMER();
if((rc=ARMCI_NbGet(&dsrc[i][0], &ddst[me][i*elems[1]], bytes,
i, &hdl_get)))
ARMCI_Error("armci_nbget failed\n",rc);
t5 += MP_TIMER()-stime; stime=MP_TIMER();
ARMCI_Wait(&hdl_get);
t6 += MP_TIMER()-stime;
}
}
}
MP_BARRIER(); ARMCI_AllFence(); MP_BARRIER();
if(VERIFY) verify_results(GET, elems);
for(i=0; i<elems[0]*elems[1]; i++) ddst[me][i]=0.0;
MP_BARRIER();
/* ------------------------ Accumulate ------------------------- */
for(i=0; i<elems[1]; i++) dsrc[me][i]=1.0; MP_BARRIER();
stride = elems[1]*sizeof(double); scale = 1.0;
for(j=0; j<ntimes; j++) {
stime=MP_TIMER();
if((rc=ARMCI_AccS(ARMCI_ACC_DBL, &scale, &dsrc[me][0], &stride,
&ddst[0][0], &stride, &bytes, 0, 0)))
ARMCI_Error("armci_acc failed\n",rc);
t7 += MP_TIMER()-stime;
MP_BARRIER(); ARMCI_AllFence(); MP_BARRIER();
if(VERIFY) verify_results(ACC, elems);
for(i=0; i<elems[0]*elems[1]; i++) ddst[me][i]=0.0;
MP_BARRIER();
}
#if 1
/* See the note below why this part is disabled */
/* ---------------------- nb-Accumulate ------------------------ */
for(i=0; i<elems[1]; i++) dsrc[me][i]=1.0; MP_BARRIER();
stride = elems[1]*sizeof(double); scale = 1.0;
for(j=0; j<ntimes; j++) {
stime=MP_TIMER();
if((rc=ARMCI_NbAccS(ARMCI_ACC_DBL, &scale, &dsrc[me][0], &stride,
&ddst[0][0], &stride, &bytes, 0, 0, &hdl_acc)))
ARMCI_Error("armci_nbacc failed\n",rc);
t8 += MP_TIMER()-stime; stime=MP_TIMER();
ARMCI_Wait(&hdl_acc);
t9 += MP_TIMER()-stime;
MP_BARRIER(); ARMCI_AllFence(); MP_BARRIER();
if(VERIFY) verify_results(ACC, elems);
for(i=0; i<elems[0]*elems[1]; i++) ddst[me][i]=0.0;
MP_BARRIER();
}
#endif
/* print timings */
if(!dry_run) if(me==0) printf("%d\t %.2e %.2e %.2e %.2e %.2e %.2e %.2e %.2e %.2e\n",
bytes, t4/ntimes, t5/ntimes, t6/ntimes, t1/ntimes,
t2/ntimes, t3/ntimes, t7/ntimes, t8/ntimes, t9/ntimes);
}
ARMCI_AllFence();
MP_BARRIER();
if(!dry_run)if(me==0){printf("O.K.\n"); fflush(stdout);}
destroy_array((void **)ddst);
destroy_array((void **)dsrc);
}