Benchmark of cudaMalloc. Allocate 1MB of memory totally with several block sizes

cudaMallocBench.cu

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

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; // 1MB

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

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

int main(int argc, char **argv)
{
    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);
}

GeForce GTX TITAN X

$ ./a.out 32
./a.out Starting...
Using Device 0: GeForce GTX TITAN X
cudaMalloc(32) x 32768 Time elapsed 0.621685 sec
cudaFree(32) x 32768 Time elapsed 0.177743 sec
$ ./a.out 64
./a.out Starting...
Using Device 0: GeForce GTX TITAN X
cudaMalloc(64) x 16384 Time elapsed 0.483772 sec
cudaFree(64) x 16384 Time elapsed 0.087232 sec
$ ./a.out 128
./a.out Starting...
Using Device 0: GeForce GTX TITAN X
cudaMalloc(128) x 8192 Time elapsed 0.398554 sec
cudaFree(128) x 8192 Time elapsed 0.048897 sec
$ ./a.out 256
./a.out Starting...
Using Device 0: GeForce GTX TITAN X
cudaMalloc(256) x 4096 Time elapsed 0.324956 sec
cudaFree(256) x 4096 Time elapsed 0.031210 sec
$ ./a.out 512
./a.out Starting...
Using Device 0: GeForce GTX TITAN X
cudaMalloc(512) x 2048 Time elapsed 0.320291 sec
cudaFree(512) x 2048 Time elapsed 0.010705 sec
$ ./a.out 1024
./a.out Starting...
Using Device 0: GeForce GTX TITAN X
cudaMalloc(1024) x 1024 Time elapsed 0.313027 sec
cudaFree(1024) x 1024 Time elapsed 0.004453 sec
9$ ./a.out 1048576
./a.out Starting...
Using Device 0: GeForce GTX TITAN X
cudaMalloc(1048576) x 1 Time elapsed 0.314088 sec
cudaFree(1048576) x 1 Time elapsed 0.000228 sec

speed: 1048576 ≒ 1024 ≒ 512 ≒ 256 > 128 > 64 > 32

NVIDIA Tesla K80 (AWS p2.xlarge)

$ ./a.out 32
./a.out Starting...
Using Device 0: Tesla K80
cudaMalloc(32) x 32768 Time elapsed 0.368449 sec
cudaFree(32) x 32768 Time elapsed 0.233109 sec
$ ./a.out 64
./a.out Starting...
Using Device 0: Tesla K80
cudaMalloc(64) x 16384 Time elapsed 0.222854 sec
cudaFree(64) x 16384 Time elapsed 0.113812 sec
$ ./a.out 128
./a.out Starting...
Using Device 0: Tesla K80
cudaMalloc(128) x 8192 Time elapsed 0.153932 sec
cudaFree(128) x 8192 Time elapsed 0.053498 sec
$ ./a.out 256
./a.out Starting...
Using Device 0: Tesla K80
cudaMalloc(256) x 4096 Time elapsed 0.122204 sec
cudaFree(256) x 4096 Time elapsed 0.034702 sec
$ ./a.out 512
./a.out Starting...
Using Device 0: Tesla K80
cudaMalloc(512) x 2048 Time elapsed 0.100751 sec
cudaFree(512) x 2048 Time elapsed 0.009651 sec
$ ./a.out 1024
./a.out Starting...
Using Device 0: Tesla K80
cudaMalloc(1024) x 1024 Time elapsed 0.096302 sec
cudaFree(1024) x 1024 Time elapsed 0.004838 sec
asag5104$ ./a.out 1048576
./a.out Starting...
Using Device 0: Tesla K80
cudaMalloc(1048576) x 1 Time elapsed 0.093849 sec
cudaFree(1048576) x 1 Time elapsed 0.000106 sec

Average for 256: 0.11
Average for 512: 0.09
Average for 1024: 0.086
Average for 2048: 0.085

speed: 1048576 ≒ 1024 ≒ 512 > 256 > 128 > 64 > 32