nvcc cudaMallocVScuMemAllocBench.cu -L /usr/local/cuda/lib64 -l cuda

cudaMallocVScuMemAllocBench.cu

#include <sys/time.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <stdio.h>

#define CHECK(call)                                                            \
{                                                                              \
    const cudaError_t error = call;                                            \
    if (error != cudaSuccess)                                                  \
    {                                                                          \
        fprintf(stderr, "Error: %s:%d, ", __FILE__, __LINE__);                 \
        fprintf(stderr, "code: %d, reason: %s\n", error,                       \
                cudaGetErrorString(error));                                    \
        exit(1);                                                               \
    }                                                                          \
}

#define CU_CHECK(call)                                                         \
{                                                                              \
    const CUresult error = call;                                               \
    if (error != CUDA_SUCCESS)                                                 \
    {                                                                          \
        fprintf(stderr, "Error: %s:%d, ", __FILE__, __LINE__);                 \
        fprintf(stderr, "code: %d\n", error);                                  \
        exit(1);                                                               \
    }                                                                          \
}

inline double seconds()
{
    struct timeval tp;
    struct timezone tzp;
    int i = gettimeofday(&tp, &tzp);
    return ((double)tp.tv_sec + (double)tp.tv_usec * 1.e-6);
}

int total_size = 1024 * 1024;

void test(int size)
{
    double iStart, iElaps;
    int num = total_size / size;
    float *d[num];
    CUdeviceptr d2[num];

    iStart = seconds();
    for (int i = 0; i < num; i++) {
      CHECK(cudaMalloc((void**)&d[i], size));
      //printf("%p\n", d[i]);
    }
    iElaps = seconds() - iStart;
    printf("cudaMalloc(%d) x %d Time elapsed %f sec\n", size, num, iElaps);

    iStart = seconds();
    for (int i = 0; i < num; i++) {
      CHECK(cudaFree(d[i]));
    }
    iElaps = seconds() - iStart;
    printf("cudaFree(%d) x %d Time elapsed %f sec\n", size, num, iElaps);

    iStart = seconds();
    for (int i = 0; i < num; i++) {
      CU_CHECK(cuMemAlloc(&d2[i], size));
      //printf("%p\n", (void*)(d2[i]));
    }
    iElaps = seconds() - iStart;
    printf("cuMemAlloc(%d) x %d Time elapsed %f sec\n", size, num, iElaps);

    iStart = seconds();
    for (int i = 0; i < num; i++) {
      CU_CHECK(cuMemFree(d2[i]));
    }
    iElaps = seconds() - iStart;
    printf("cuMemFree(%d) x %d Time elapsed %f sec\n", size, num, iElaps);
}

int main(int argc, char **argv)
{
     if (argc < 2) {
        printf("%s size\n", argv[0]);
        exit(1);
    }
    printf("%s Starting...\n", argv[0]);

    // set up device
    int dev = 0;
    cudaDeviceProp deviceProp;
    cudaGetDeviceProperties(&deviceProp, dev);
    printf("Using Device %d: %s\n", dev, deviceProp.name);
    cudaSetDevice(dev);

    int size = atoi(argv[1]);
    test(size);

    return(0);
}

cuMemAlloc looked faster, but it was because cudaMalloc creates a context on a first call.