NASP/cuda-samples
35
0
Fork 0
mirror of https://github.com/NVIDIA/cuda-samples.git synced 2024-11-24 19:49:19 +08:00
Code Issues Packages Projects Releases Wiki Activity

[p2pBandwidthLatency] increase default buffer size and add support to specific buffer size through command line

Browse Source
This commit is contained in:
Mahesh Doijade 2020-06-30 19:05:55 +05:30
parent 0ec4bd58e5
commit e6ce58fef4
1 changed files with 15 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
Samples/p2pBandwidthLatencyTest/p2pBandwidthLatencyTest.cu
View File

@ -99,7 +99,8 @@ void printHelp(void) {
printf( printf(
"--p2p_read\tUse P2P reads for data transfers between GPU pairs and show " "--p2p_read\tUse P2P reads for data transfers between GPU pairs and show "
"corresponding results.\n \t\tDefault used is P2P write operation.\n"); "corresponding results.\n \t\tDefault used is P2P write operation.\n");
printf("--sm_copy\t\tUse SM intiated p2p transfers instead of Copy Engine\n"); printf("--sm_copy Use SM intiated p2p transfers instead of Copy Engine\n");
printf("--numElems=<NUM_OF_INT_ELEMS> Number of integer elements to be used in p2p copy.\n");
} }
void checkP2Paccess(int numGPUs) { void checkP2Paccess(int numGPUs) {
@ -145,8 +146,7 @@ void performP2PCopy(int *dest, int destDevice, int *src, int srcDevice,
} }
} }
void outputBandwidthMatrix(int numGPUs, bool p2p, P2PDataTransfer p2p_method) { void outputBandwidthMatrix(int numElems, int numGPUs, bool p2p, P2PDataTransfer p2p_method) {
int numElems = 10000000;
int repeat = 5; int repeat = 5;
volatile int *flag = NULL; volatile int *flag = NULL;
vector<int *> buffers(numGPUs); vector<int *> buffers(numGPUs);
@ -287,8 +287,7 @@ void outputBandwidthMatrix(int numGPUs, bool p2p, P2PDataTransfer p2p_method) {
cudaCheckError(); cudaCheckError();
} }
void outputBidirectionalBandwidthMatrix(int numGPUs, bool p2p) { void outputBidirectionalBandwidthMatrix(int numElems, int numGPUs, bool p2p) {
int numElems = 10000000;
int repeat = 5; int repeat = 5;
volatile int *flag = NULL; volatile int *flag = NULL;
vector<int *> buffers(numGPUs); vector<int *> buffers(numGPUs);
@ -604,7 +603,7 @@ void outputLatencyMatrix(int numGPUs, bool p2p, P2PDataTransfer p2p_method) {
} }
int main(int argc, char **argv) { int main(int argc, char **argv) {
int numGPUs; int numGPUs, numElems = 40000000;
P2PDataTransfer p2p_method = P2P_WRITE; P2PDataTransfer p2p_method = P2P_WRITE;
cudaGetDeviceCount(&numGPUs); cudaGetDeviceCount(&numGPUs);
@ -624,6 +623,11 @@ int main(int argc, char **argv) {
p2p_mechanism = SM; p2p_mechanism = SM;
} }
// number of elements of int to be used in copy.
if (checkCmdLineFlag(argc, (const char **)argv, "numElems")) {
numElems = getCmdLineArgumentInt(argc, (const char **)argv, "numElems");
}
printf("[%s]\n", sSampleName); printf("[%s]\n", sSampleName);
// output devices // output devices
@ -662,17 +666,17 @@ int main(int argc, char **argv) {
} }
printf("Unidirectional P2P=Disabled Bandwidth Matrix (GB/s)\n"); printf("Unidirectional P2P=Disabled Bandwidth Matrix (GB/s)\n");
outputBandwidthMatrix(numGPUs, false, P2P_WRITE); outputBandwidthMatrix(numElems, numGPUs, false, P2P_WRITE);
printf("Unidirectional P2P=Enabled Bandwidth (P2P Writes) Matrix (GB/s)\n"); printf("Unidirectional P2P=Enabled Bandwidth (P2P Writes) Matrix (GB/s)\n");
outputBandwidthMatrix(numGPUs, true, P2P_WRITE); outputBandwidthMatrix(numElems, numGPUs, true, P2P_WRITE);
if (p2p_method == P2P_READ) { if (p2p_method == P2P_READ) {
printf("Unidirectional P2P=Enabled Bandwidth (P2P Reads) Matrix (GB/s)\n"); printf("Unidirectional P2P=Enabled Bandwidth (P2P Reads) Matrix (GB/s)\n");
outputBandwidthMatrix(numGPUs, true, p2p_method); outputBandwidthMatrix(numElems, numGPUs, true, p2p_method);
} }
printf("Bidirectional P2P=Disabled Bandwidth Matrix (GB/s)\n"); printf("Bidirectional P2P=Disabled Bandwidth Matrix (GB/s)\n");
outputBidirectionalBandwidthMatrix(numGPUs, false); outputBidirectionalBandwidthMatrix(numElems, numGPUs, false);
printf("Bidirectional P2P=Enabled Bandwidth Matrix (GB/s)\n"); printf("Bidirectional P2P=Enabled Bandwidth Matrix (GB/s)\n");
outputBidirectionalBandwidthMatrix(numGPUs, true); outputBidirectionalBandwidthMatrix(numElems, numGPUs, true);
printf("P2P=Disabled Latency Matrix (us)\n"); printf("P2P=Disabled Latency Matrix (us)\n");
outputLatencyMatrix(numGPUs, false, P2P_WRITE); outputLatencyMatrix(numGPUs, false, P2P_WRITE);