$ cat t21.cu
#include <stdio.h>
#include <time.h>
#include <sys/time.h>
#define DSIZE 32768
#define NUM_COPY 10
int main(){

  struct timeval t1, t2;
  int *d_A, *h_A;
  int copy_count = NUM_COPY;
  cudaHostAlloc((void **)&h_A, DSIZE*sizeof(int), cudaHostAllocDefault);
  cudaMalloc((void **)&d_A, DSIZE*sizeof(int));
  cudaStream_t my_stream;
  cudaStreamCreate(&my_stream);
  cudaMemcpyAsync(d_A, h_A, DSIZE*sizeof(int), cudaMemcpyHostToDevice, my_stream);
  gettimeofday(&t1, NULL);
  gettimeofday(&t2, NULL);
  int time = (t2.tv_sec * (unsigned int)1e6 + t2.tv_usec) - (t1.tv_sec * (unsigned int)1e6 + t1.tv_usec);
  printf("null timing = %dus\n", time);
  for (int i = 0; i < 5; i ++){
    gettimeofday(&t1, NULL);
    for (int j = 0; j < copy_count; j++)
      cudaMemcpyAsync(d_A, h_A, DSIZE*sizeof(int), cudaMemcpyHostToDevice, my_stream);
    gettimeofday(&t2, NULL);
    time = (t2.tv_sec * (unsigned int)1e6 + t2.tv_usec) - (t1.tv_sec * (unsigned int)1e6 + t1.tv_usec);
    printf("%d call timing = %fus average\n", copy_count, (float)time/(float)copy_count);
    copy_count *= NUM_COPY;
    }
  return 0;
}


$ nvcc -arch=sm_20 -o t21 t21.cu
$ ./t21
null timing = 1us
10 call timing = 2.800000us average
100 call timing = 2.410000us average
1000 call timing = 2.409000us average
10000 call timing = 26.297800us average
100000 call timing = 26.338289us average
$