int main(int argc, char *argv[])
{
double minTime = 99999.0;
double testTime = 0.0;
double maxBidirBw = 0.0;
double maxUnidirBw = 0.0;
int maxBidirSize = 0;
int maxUnidirSize = 0;
char* procFile = NULL;
int* procList = NULL;
int procListSize;
int rank, wsize, iters, procs;
int messStart, messStop, messFactor, messSize;
int procsPerNode, totalops, procIdx;
int useBarrier, printPairs, useNearestRank, dummy;
double currBw;
char allocPattern;
MPI_Comm activeComm;
command = argv[0];
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &wsize);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
if(rank == 0) {
//sleep(5);
//while (!__sync_bool_compare_and_swap(&r_sync, 59, 0)){ sleep(2); printf("sync value %d \n", r_sync); };
int k = 0;
#pragma omp parallel
#pragma omp atomic
k++;
printf ("Done Sync. Number of Threads counted = %i\n",k);
int N = 1048576;
int i = 0;
int j = 0;
double start, diff;
for(j=0; j < 50 ; j++) {
#pragma omp parallel for
for (i=0; i<N; i++) {
a[i] = 1.0;
b[i] = 2.0;
}
start = MPI_Wtime();
#pragma omp parallel for
for (i = 0; i < N; i++) {
a[i]= b[i];
}
diff = MPI_Wtime() - start ;
printf("ELAPSED time to copy : %11.6f N: %d \n",diff, N);
}
}else{
//__sync_fetch_and_add(&r_sync, 1);
printf("OK other ranks \n");
//sleep(100);
}
MPI_Barrier(MPI_COMM_WORLD);
MPI_Finalize();
return 0 ;
if ( !processArgs(argc, argv, rank, wsize, &iters, &dummy,
&messStart, &messStop, &messFactor,
&procFile, &procsPerNode, &allocPattern,
&printPairs, &useBarrier, &useNearestRank) )
{
if ( rank == 0 )
printUse();
MPI_Finalize();
exit(-1);
}
if ( ! getProcList(procFile, wsize, &procList, &procListSize,
procsPerNode, allocPattern) )
{
if ( procFile )
printf("Failed to get process list from file %s.\n", procFile);
else
printf("Failed to allocate process list.\n");
exit(-1);
}
if ( rank == 0 )
printReportHeader();
//Set up intra-node shared memory structures.
if(rank == 0) {
//One rank per node allocates shared send request lists.
req_array = malloc(wsize*sizeof(ReqInfo));
/*if(!IS_SM_BUF((void*)req_array)){
printf("Failed to allocate from SM region :%p \n", req_array);
exit(-1);
}*/
}
//broadcast address of the pointer variable so that each rank has access to it
MPI_Bcast(&req_array , 1, MPI_LONG, 0, MPI_COMM_WORLD);
//printf("Broadcasting of shared mem region complete! rank : %d region_addr : %lu region_ptr : %lu \n", rank, &req_array, req_array);
//printf("Access shm region ! rank : %d region for sync : %lu \n", rank, req_array[rank].sync);
/*
if(rank == 0){
char* comBuf = (char*)malloc(10);
memset(comBuf, 1, 10);
req_array[30].buffer = comBuf ;
comBuf[0] = 'n';
req_array[30].sync = malloc(sizeof(int));
req_array[30].sync[0] = 12;
printf("rank : %d done sending buff ptr : %p sync ptr : %p \n",rank, comBuf, req_array[30].sync );
printf("sleeping ! pid() %d \n", getpid());
sleep(40000);
}else if(rank == 30){
while(req_array[30].sync == NULL){
}
while(req_array[30].sync[0] != 12){
}
printf("rank : %d buffer value: %c sync value : %d buff ptr : %p sync ptr : %p \n",
rank, req_array[30].buffer[0], req_array[30].sync[0], req_array[30].buffer, req_array[30].sync );
printf("sleeping ! pid() %d \n", getpid());
sleep(40000);
}
MPI_Barrier(MPI_COMM_WORLD);
return 0;*/
printPairs = 1 ;
for ( procIdx = 0; procIdx < procListSize; procIdx++ )
{
procs = procList[procIdx];
if ( rank == 0 && printPairs )
{
printActivePairs(procs, rank, wsize, procsPerNode,
allocPattern, useNearestRank);
}
/* Create Communicator of all active processes */
createActiveComm(procs, rank, wsize, procsPerNode,
allocPattern, printPairs, useNearestRank, &activeComm);
// messStart = 8388608;
// messStop = 4096 ;
// messStop = 8388608 ;
for ( messSize = messStart; messSize <= messStop;
messSize *= messFactor )
{
testTime = runUnicomTest(procs, messSize, iters, rank,
wsize, procsPerNode, allocPattern,
useBarrier, useNearestRank, &activeComm);
if ( rank == 0 && testTime > 0 )
{
totalops = iters * (procs/2);
currBw = ((double)totalops*(double)messSize/testTime)/1000000;
if ( currBw > maxUnidirBw )
{
maxUnidirBw = currBw;
maxUnidirSize = messSize;
}
if ( testTime < minTime )
minTime = testTime;
}
}
if ( activeComm != MPI_COMM_NULL )
MPI_Comm_free(&activeComm);
if ( rank == 0 )
printf("\n");
}
for ( procIdx = 0; procIdx < procListSize; procIdx++ )
{
procs = procList[procIdx];
if ( rank == 0 && printPairs )
{
printActivePairs(procs, rank, wsize, procsPerNode,
allocPattern, useNearestRank);
}
// Create Communicator of all active processes
createActiveComm(procs, rank, wsize, procsPerNode,
allocPattern, printPairs, useNearestRank, &activeComm);
for ( messSize = messStart; messSize <= messStop;
messSize *= messFactor )
{
testTime = runBicomTest(procs, messSize, iters, rank,
wsize, procsPerNode, allocPattern,
useBarrier, useNearestRank, &activeComm);
if ( rank == 0 && testTime > 0 )
{
totalops = iters * procs;
currBw = (((double)totalops*(double)messSize)/testTime)/1000000;
if ( currBw > maxBidirBw )
{
maxBidirBw = currBw;
maxBidirSize = messSize;
}
if ( testTime < minTime )
minTime = testTime;
}
}
if ( activeComm != MPI_COMM_NULL )
MPI_Comm_free(&activeComm);
if ( rank == 0 )
printf("\n");
}
if ( rank == 0 )
{
printf("Max Unidirectional Bandwith : %13.2f for message size of %7d bytes\n",
maxUnidirBw, maxUnidirSize);
printf("Max Bidirectional Bandwith : %13.2f for message size of %7d bytes\n",
maxBidirBw, maxBidirSize);
printParameters(iters, messStart, messStop, messFactor,
procFile, procsPerNode, allocPattern, useBarrier);
free(req_array);
}
printReportFooter(minTime, rank, wsize, procsPerNode, useNearestRank);
free(procList);
MPI_Finalize();
exit(0);
}
void
runTest ( TESTPARAMS * testParams )
{
int procIdx, procs, messIdx;
MPI_Comm activeComm = MPI_COMM_NULL;
int *messList, messListSize;
int testCount, iters;
unsigned int i;
char buf[256];
int width;
double result;
STATSTRUCT dp, sum;
prestaRankPrint ( 0,
"\n\n%s Test Results \n(tasks, size, ops/sample, samples) : min/mean/max \n",
testParams->name );
if ( argStruct.verbose )
{
if ( testParams->id == LATEN )
{
if ( argStruct.sumLocalBW == 1 )
{
sprintf ( buf, outputCharLBWFormat, "Test", "Processes",
"Op Size (bytes)", "Ops", "Latency (us)" );
}
else
{
sprintf ( buf, outputCharFormat, "Test", "Processes",
"Op Size (bytes)", "Ops", "BW (MB)", "Latency (us)" );
}
}
else
{
if ( argStruct.sumLocalBW == 1 )
{
sprintf ( buf, outputCharLBWFormat, "Test", "Processes",
"Op Size (bytes)", "Ops", "BW (MB)" );
}
else
{
sprintf ( buf, outputCharFormat, "Test", "Processes",
"Op Size (bytes)", "Ops", "BW (MB)", "Op time (us)" );
}
}
prestaRankPrint ( 0, "%s", buf );
width = strlen ( buf );
}
else
width = 80;
for ( i = 0; i < width; i++ )
buf[i] = '-';
buf[i] = '\0';
prestaRankPrint ( 0, "%s\n", buf );
messList = testParams->messList;
messListSize = testParams->messListSize;
testParams->maxBW = 0.0;
testParams->maxBWMessSize = 0;
init_stats ( &sum, 0, 0, 0, 0 );
for ( procIdx = 0; procIdx < argStruct.procListSize; procIdx++ )
{
procs = argStruct.procList[procIdx];
if ( procs > wsize )
procs = wsize;
/* Create Communicator of all active processes */
procs = createActiveComm ( procs, argStruct.procsPerNode,
argStruct.allocPattern,
argStruct.useNearestRank, &activeComm );
prestaDebug ( "rank %d returned from createActiveCom\n", rank );
prestaDebug ( "messListSize is %d \n", messListSize );
for ( messIdx = 0; messIdx < messListSize; messIdx++ )
{
if ( argStruct.iterList != NULL && argStruct.iterList[messIdx] != 0 )
iters = argStruct.iterList[messIdx];
else
iters = argStruct.iters;
if ( argStruct.testCountList != NULL )
{
fprintf ( stderr,
"Before init_stats! messIdx is %d, procIdx is %d, testCountList is %p\n",
messIdx, procIdx, argStruct.testCountList );
init_stats ( &dp, procs, messList[messIdx], iters,
argStruct.testCountList[procIdx] );
}
else
{
init_stats ( &dp, procs, messList[messIdx], iters, argStruct.samples );
}
for ( testCount = 0; testCount < dp.samples; testCount++ )
{
/* Run test and save current result */
testParams->rankResult =
testParams->testFunc ( dp.msize, iters, &activeComm );
/* TODO : Remove this if unnecessary */
if ( testParams->rankResult < minTime )
minTime = testParams->rankResult;
if ( !generateResults
( testParams, procs, dp.msize, iters, &result ) )
prestaAbort ( "Failed to generate test results." );
update_stats ( &dp, result );
update_stats ( &sum, result );
}
if ( testParams->id == LATEN )
{
prestaRankPrint ( 0,
"(%6d, %9d, %6d, %6d): %6.3f / %6.3f / %6.3f\n",
dp.tasks, dp.msize, dp.iters, dp.samples, dp.min,
dp.mean, dp.max );
}
else
{
prestaRankPrint ( 0,
"(%6d, %7d, %5d, %5d): %12.3f / %12.3f / %12.3f\n",
dp.tasks, dp.msize, dp.iters, dp.samples, dp.min,
dp.mean, dp.max );
}
}
}
if ( testParams->id == LATEN )
{
prestaRankPrint ( 0,
"\nSummary : min/mean/max = %6.3f / %6.3f / %6.3f\n",
sum.min, sum.mean, sum.max );
}
else
{
prestaRankPrint ( 0,
"\nSummary : min/mean/max = %12.3f / %12.3f / %12.3f\n",
sum.min, sum.mean, sum.max );
}
if ( rank == 0 && argStruct.printPairs )
{
printActivePairs ( procs, argStruct.procsPerNode,
argStruct.allocPattern, argStruct.useNearestRank );
}
if ( activeComm != MPI_COMM_NULL )
MPI_Comm_free ( &activeComm );
}
main(int argc, char** argv)
{
int rank, wsize, iters, i, procs, currtarg, dummy;
double diff = 0.0;
double start, max, mintime = 9999;
MPI_Status stat;
char comBuf;
MPI_Comm activeComm;
char* procFile = NULL;
int* procList = NULL;
int procListSize;
int messStart, messStop, messFactor;
int procsPerNode, procIdx, useBarrier, printPairs, useNearestRank;
char allocPattern;
command = argv[0];
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &wsize);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
if ( !processArgs(argc, argv, rank, wsize, &iters,
&dummy, &messStart, &messStop, &messFactor,
&procFile, &procsPerNode, &allocPattern,
&printPairs, &useBarrier, &useNearestRank) )
{
if ( rank == 0 )
printUse();
MPI_Finalize();
exit(-1);
}
if ( ! getProcList(procFile, wsize, &procList, &procListSize,
procsPerNode, allocPattern) )
{
if ( procFile )
printf("Failed to get process list from file %s.\n", procFile);
else
printf("Failed to allocate process list.\n");
exit(-1);
}
if ( rank == 0 )
printReportHeader();
currtarg = getTargetRank(rank, wsize, procsPerNode, useNearestRank);
for ( procIdx = 0; procIdx < procListSize; procIdx++ )
{
procs = procList[procIdx];
if ( printPairs )
{
printActivePairs(procs, rank, wsize, procsPerNode,
allocPattern, useNearestRank);
}
/* Create Communicator of all active processes */
createActiveComm(procs, rank, wsize, procsPerNode,
allocPattern, printPairs, useNearestRank, &activeComm);
if ( isActiveProc(rank, wsize, procsPerNode, procs,
allocPattern, useNearestRank) )
{
if ( rank < currtarg )
{
/* Ensure pair communication has been initialized */
MPI_Recv(&comBuf, 0, MPI_INT, currtarg, 0, MPI_COMM_WORLD, &stat);
MPI_Send(&comBuf, 0, MPI_INT, currtarg, 0, MPI_COMM_WORLD);
}
else
{
/* Ensure pair communication has been initialized */
MPI_Send(&comBuf, 0, MPI_INT, currtarg, 0, MPI_COMM_WORLD);
MPI_Recv(&comBuf, 0, MPI_INT, currtarg, 0, MPI_COMM_WORLD, &stat);
}
//generic_barrier(activeComm);
MPI_Barrier(activeComm);
//generic_barrier(activeComm);
MPI_Barrier(activeComm);
if ( rank < currtarg )
{
/* Time operation loop */
start = MPI_Wtime();
for ( i = 0; i < iters; i++ )
{
MPI_Send(&comBuf, 0, MPI_INT, currtarg, 0, MPI_COMM_WORLD);
MPI_Recv(&comBuf, 0, MPI_INT, currtarg, 0, MPI_COMM_WORLD, &stat);
}
}
else
{
/* Time operation loop */
start = MPI_Wtime();
for ( i = 0; i < iters; i++ )
{
MPI_Recv(&comBuf, 0, MPI_INT, currtarg, 0, MPI_COMM_WORLD, &stat);
MPI_Send(&comBuf, 0, MPI_INT, currtarg, 0, MPI_COMM_WORLD);
}
}
if ( useBarrier )
MPI_Barrier(activeComm);
//generic_barrier(activeComm);
diff = MPI_Wtime() - start;
}
if ( activeComm != MPI_COMM_NULL )
MPI_Comm_free(&activeComm);
/* Get maximum sample length */
MPI_Reduce(&diff, &max, 1, MPI_DOUBLE, MPI_MAX, 0, MPI_COMM_WORLD);
if ( rank == 0 )
{
if ( max < mintime )
mintime = max;
printf(outputFormat, procs, max/iters/2*1000000);
}
}
if ( rank == 0 )
{
printParameters(iters, procFile, procsPerNode,
allocPattern, useBarrier);
}
printReportFooter(mintime, rank, wsize, procsPerNode, useNearestRank);
MPI_Finalize();
exit(0);
}
int main(int argc, char *argv[])
{
double minTime = 99999.0;
double testTime = 0.0;
double maxBidirBw = 0.0;
double maxUnidirBw = 0.0;
int maxBidirSize = 0;
int maxUnidirSize = 0;
char* procFile = NULL;
int* procList = NULL;
int procListSize;
int rank, wsize, iters, procs;
int messStart, messStop, messFactor, messSize;
int procsPerNode, totalops, procIdx;
int useBarrier, printPairs, useNearestRank, dummy;
double currBw;
char allocPattern;
MPI_Comm activeComm;
command = argv[0];
MPI_Init(&argc, &argv);
MPI_Comm_size(MPI_COMM_WORLD, &wsize);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
if ( !processArgs(argc, argv, rank, wsize, &iters, &dummy,
&messStart, &messStop, &messFactor,
&procFile, &procsPerNode, &allocPattern,
&printPairs, &useBarrier, &useNearestRank) )
{
if ( rank == 0 )
printUse();
MPI_Finalize();
exit(-1);
}
if ( ! getProcList(procFile, wsize, &procList, &procListSize,
procsPerNode, allocPattern) )
{
if ( procFile )
printf("Failed to get process list from file %s.\n", procFile);
else
printf("Failed to allocate process list.\n");
exit(-1);
}
if ( rank == 0 )
printReportHeader();
for ( procIdx = 0; procIdx < procListSize; procIdx++ )
{
procs = procList[procIdx];
if ( rank == 0 && printPairs )
{
printActivePairs(procs, rank, wsize, procsPerNode,
allocPattern, useNearestRank);
}
/* Create Communicator of all active processes */
createActiveComm(procs, rank, wsize, procsPerNode,
allocPattern, printPairs, useNearestRank, &activeComm);
for ( messSize = messStart; messSize <= messStop;
messSize *= messFactor )
{
testTime = runUnicomTest(procs, messSize, iters, rank,
wsize, procsPerNode, allocPattern,
useBarrier, useNearestRank, &activeComm);
if ( rank == 0 && testTime > 0 )
{
totalops = iters * (procs/2);
currBw = ((double)totalops*(double)messSize/testTime)/1000000;
if ( currBw > maxUnidirBw )
{
maxUnidirBw = currBw;
maxUnidirSize = messSize;
}
if ( testTime < minTime )
minTime = testTime;
}
}
if ( activeComm != MPI_COMM_NULL )
MPI_Comm_free(&activeComm);
if ( rank == 0 )
printf("\n");
}
for ( procIdx = 0; procIdx < procListSize; procIdx++ )
{
procs = procList[procIdx];
if ( rank == 0 && printPairs )
{
printActivePairs(procs, rank, wsize, procsPerNode,
allocPattern, useNearestRank);
}
/* Create Communicator of all active processes */
createActiveComm(procs, rank, wsize, procsPerNode,
allocPattern, printPairs, useNearestRank, &activeComm);
for ( messSize = messStart; messSize <= messStop;
messSize *= messFactor )
{
testTime = runBicomTest(procs, messSize, iters, rank,
wsize, procsPerNode, allocPattern,
useBarrier, useNearestRank, &activeComm);
if ( rank == 0 && testTime > 0 )
{
totalops = iters * procs;
currBw = (((double)totalops*(double)messSize)/testTime)/1000000;
if ( currBw > maxBidirBw )
{
maxBidirBw = currBw;
maxBidirSize = messSize;
}
if ( testTime < minTime )
minTime = testTime;
}
}
if ( activeComm != MPI_COMM_NULL )
MPI_Comm_free(&activeComm);
if ( rank == 0 )
printf("\n");
}
if ( rank == 0 )
{
printf("Max Unidirectional Bandwith : %13.2f for message size of %7d bytes\n",
maxUnidirBw, maxUnidirSize);
printf("Max Bidirectional Bandwith : %13.2f for message size of %7d bytes\n",
maxBidirBw, maxBidirSize);
printParameters(iters, messStart, messStop, messFactor,
procFile, procsPerNode, allocPattern, useBarrier);
}
printReportFooter(minTime, rank, wsize, procsPerNode, useNearestRank);
free(procList);
MPI_Finalize();
exit(0);
}
