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);
}
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);
}
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 lu(int n, int bs, int me)
{
int i, il, j, jl, k, kl;
int I, J, K;
double *A, *B, *C, *D;
int dimI, dimJ, dimK;
int strI, strJ, strK;
unsigned int t1, t2, t3, t4, t11, t22;
int diagowner, destp, hc, m;
double *dbuf;
armci_hdl_t handle[2*MAXPROC];
int saved[MAXPROC];
dbuf = (double *)ARMCI_Malloc_local((armci_size_t) block_size*block_size*sizeof(double));
for (k=0, K=0; k<n; k+=bs, K++) {
kl = k + bs;
if (kl > n) {
kl = n;
strK = kl - k;
} else {
strK = bs;
}
/* factor diagonal block */
diagowner = block_owner(K, K);
if (diagowner == me) {
A = a[K+K*nblocks];
lu0(A, strK, strK); /* impl algo on this diag block */
}
MP_BARRIER();
/* divide column k by diagonal block */
if(block_owner(K, K) == me)
D = a[K+K*nblocks];
else {
D = dbuf;
get_remote(D, K, K);
}
for (i=kl, I=K+1; i<n; i+=bs, I++) {
if (block_owner(I, K) == me) { /* parcel out blocks */
il = i + bs;
if (il > n) {
il = n;
strI = il - i;
} else {
strI = bs;
}
A = a[I+K*nblocks];
bdiv(A, D, strI, strK, strI, strK);
/* Pre-put this block to the block-owners of all blocks on the I-th row with a non-blocking put*/
memset (saved, 0, sizeof(saved));
for (m = K+1; m < nblocks; m++) {
destp = block_owner (I, m);
if (destp != me && !saved[destp]) {
ARMCI_NbPut(A, bufc[destp*nblocks + I], strI*strK*sizeof(double), destp, NULL);
saved[destp] = 1;
}
}
}
} /* end of for (i=k1, I=K+1...) */
/* modify row k by diagonal block */
for (j=kl, J=K+1; j<n; j+=bs, J++) {
if (block_owner(K, J) == me) { /* parcel out blocks */
jl = j+bs;
if (jl > n) {
jl = n;
strJ = jl - j;
} else {
strJ = bs;
}
A = a[K+J*nblocks];
bmodd(D, A, strK, strJ, strK, strK);
/* Pre-put this block to the block-owners of all blocks on the J-th column with a non-blocking put*/
memset (saved, 0, sizeof(saved));
for (m = K+1; m < nblocks; m++) {
destp = block_owner (m, J);
if (destp != me && !saved[destp]) {
ARMCI_NbPut(A, bufr[destp*nblocks + J], strK*strJ*sizeof(double), destp, NULL);
saved[destp] = 1;
}
}
}
}
ARMCI_WaitAll();
ARMCI_AllFence();
MP_BARRIER();
/* modify subsequent block columns */
for (i=kl, I=K+1; i<n; i+=bs, I++) {
il = i+bs;
if (il > n) {
il = n;
strI = il - i;
} else {
strI = bs;
}
for (j=kl, J=K+1; j<n; j+=bs, J++) {
jl = j + bs;
if (jl > n) {
jl = n;
strJ= jl - j;
} else {
strJ = bs;
}
if (block_owner(I, J) == me) { /* parcel out blocks */
if(block_owner(I,K) == me)
A = a[I+K*nblocks];
else {
A = bufc[me*nblocks+I];
}
if(block_owner(K,J) == me)
B = a[K+J*nblocks];
else
B = bufr[me*nblocks + J];
C = a[I+J*nblocks];
bmod(A, B, C, strI, strJ, strK, strI, strK, strI);
}
}
}
}
ARMCI_Free_local(dbuf);
}
