cudaMemcpy bandwidth test

cudaMemcpyBandwidthTest.cu

#include <iostream>
#include <ctime>
#include <cuda_runtime.h>

#define MEGA_BYTE (1'000'000)

const int N = 100 * MEGA_BYTE; // Size of data
const int numIterations = 100; // iteration

// CUDA error check macro
#define CUDA_CHECK(call) \
do { \
    cudaError_t err = call; \
    if (err != cudaSuccess) { \
        std::cerr << "CUDA Error: " << cudaGetErrorString(err) << " (" << err << ") at " << __FILE__ << ":" << __LINE__ << std::endl; \
        exit(err); \
    } \
} while (0)

int main() {
    // allocate memory on host and device
    int* h_data = new int[N];
    int* d_data;
    CUDA_CHECK(cudaMalloc((void**)&d_data, N * sizeof(int)));

    // data initialization
    for (int i = 0; i < N; i++) {
        h_data[i] = i;
    }

    // measure time for dev to host
    clock_t start, end;
    double devToHostTime = 0.0;

    for (int iter = 0; iter < numIterations; iter++) {
        start = clock();
        CUDA_CHECK(cudaMemcpy(h_data, d_data, N * sizeof(int), cudaMemcpyDeviceToHost));
        end = clock();
        devToHostTime += (double)(end - start) / CLOCKS_PER_SEC;
    }

    // measure time for host to dev
    double hostToDevTime = 0.0;

    for (int iter = 0; iter < numIterations; iter++) {
        start = clock();
        CUDA_CHECK(cudaMemcpy(d_data, h_data, N * sizeof(int), cudaMemcpyHostToDevice));
        end = clock();
        hostToDevTime += (double)(end - start) / CLOCKS_PER_SEC;
    }

    // convert to average
    devToHostTime /= numIterations;
    hostToDevTime /= numIterations;

    // show result
    std::cout << "Average bandwidth for Device to Host memory copy: " << N * sizeof(int) / devToHostTime / MEGA_BYTE << " Mbytes / s" << std::endl;
    std::cout << "Average bandwidth for Host to Device memory copy: " << N * sizeof(int) / hostToDevTime / MEGA_BYTE << " Mbytes / s" << std::endl;

    // free memory
    delete[] h_data;
    CUDA_CHECK(cudaFree(d_data));

    return 0;
}

MDX (A100 with vmware virtualization)

$ nvidia-smi
Sun Sep 17 20:11:06 2023       
+---------------------------------------------------------------------------------------+
| NVIDIA-SMI 535.104.05             Driver Version: 535.104.05   CUDA Version: 12.2     |
|-----------------------------------------+----------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |         Memory-Usage | GPU-Util  Compute M. |
|                                         |                      |               MIG M. |
|=========================================+======================+======================|
|   0  NVIDIA A100-SXM4-40GB          On  | 00000000:0C:00.0 Off |                    0 |
| N/A   26C    P0              45W / 400W |      4MiB / 40960MiB |      0%      Default |
|                                         |                      |             Disabled |
+-----------------------------------------+----------------------+----------------------+
                                                                                         
+---------------------------------------------------------------------------------------+
| Processes:                                                                            |
|  GPU   GI   CI        PID   Type   Process name                            GPU Memory |
|        ID   ID                                                             Usage      |
|=======================================================================================|
|  No running processes found                                                           |
+---------------------------------------------------------------------------------------+

Output

Average bandwidth for Device to Host memory copy: 9817.57 Mbytes / s
Average bandwidth for Host to Device memory copy: 10713.9 Mbytes / s