PCI-e Transactions in UVA

Question

The memory copy calls from CPU-GPU and vice-versa in Unified Virtual Addressing (UVA) in CUDA are scheduled internally. However, the nvprof cuda profiler does not report the PCI-e bus transactions for UVA. Is there any way to know the data transfers that are taking place from host to device and device to host?

Answer 1

Yes, it's possible to get nvprof to report on Unified Memory activities. You may wish to study the options that are available using

nvprof --help

If you combine the --print-gpu-trace and --unified-memory-profiling per-process-device options, you should get some results indicating the UM activity.

Here is an example:

$ cat t476.cu
#include <stdio.h>
#define cudaCheckErrors(msg) \
    do { \
        cudaError_t __err = cudaGetLastError(); \
        if (__err != cudaSuccess) { \
            fprintf(stderr, "Fatal error: %s (%s at %s:%d)\n", \
                msg, cudaGetErrorString(__err), \
                __FILE__, __LINE__); \
            fprintf(stderr, "*** FAILED - ABORTING\n"); \
            exit(1); \
        } \
    } while (0)

__global__ void mykernel(int *d_data){

  printf("Data = %d\n", *d_data);
  *d_data = 0;
}

int main(){

  cudaDeviceProp myprop;
  int mydevice;
  int numdevices;
  cudaGetDeviceCount(&numdevices);
  cudaCheckErrors("get dev count fail");
  for (mydevice = 0; mydevice < numdevices; mydevice++){
    cudaGetDeviceProperties(&myprop, mydevice);
    printf("device %d: %s\n", mydevice, myprop.name);
    printf("device %d supports unified addressing: ", mydevice);
    if (myprop.unifiedAddressing) printf(" yes\n");
    else printf("  no\n");
    printf("device %d supports managed memory: ", mydevice);
    if (myprop.managedMemory) printf(" yes\n");
    else printf("  no\n");
    }
  cudaSetDevice(--mydevice);
  printf("using device %d\n", mydevice);
  int h_data = 1;
  int *d_data;
  cudaMalloc(&d_data, sizeof(int));
  cudaMemcpy(d_data, &h_data, sizeof(int), cudaMemcpyHostToDevice);
  mykernel<<<1,1>>>(d_data);
  cudaMemcpy(&h_data, d_data, sizeof(int), cudaMemcpyDeviceToHost);
  printf("data = %d\n", h_data);
  printf("now testing managed memory\n");
  int *m_data;
  cudaMallocManaged(&m_data, sizeof(int));
  cudaCheckErrors("managed mem fail");
  *m_data = 1;
  mykernel<<<1,1>>>(m_data);
  cudaDeviceSynchronize();
  printf("data = %d\n", m_data);
  cudaCheckErrors("some error");
  return 0;
}
$ nvcc -arch=sm_35 -o t476 t476.cu                                                                             
$ nvprof --print-gpu-trace --unified-memory-profiling per-process-device ./t476
==5114== NVPROF is profiling process 5114, command: ./t476
device 0: GeForce GT 640
device 0 supports unified addressing:  yes
device 0 supports managed memory:  yes
using device 0
Data = 1
data = 0
now testing managed memory
Data = 1
data = 0
==5114== Profiling application: ./t476
==5114== Profiling result:
   Start  Duration            Grid Size      Block Size     Regs*    SSMem*    DSMem*      Size  Throughput           Device   Context    Stream             Unified Memory  Name
1.10622s  1.1200us                    -               -         -         -         -        4B  3.5714MB/s  GeForce GT 640          1         7                          -  [CUDA memcpy HtoD]
1.10687s  64.481us              (1 1 1)         (1 1 1)        32        0B        0B         -           -  GeForce GT 640          1         7                          -  mykernel(int*) [102]
1.10693s  2.3360us                    -               -         -         -         -        4B  1.7123MB/s  GeForce GT 640          1         7                          -  [CUDA memcpy DtoH]
1.12579s         -                    -               -         -         -         -         -           -  GeForce GT 640          -         -                          0  [Unified Memory CPU page faults]
1.12579s         -                    -               -         -         -         -         -           -  GeForce GT 640          -         -                        0 B  [Unified Memory Memcpy DtoH]
1.12579s         -                    -               -         -         -         -         -           -  GeForce GT 640          -         -                        0 B  [Unified Memory Memcpy HtoD]
1.12590s  64.097us              (1 1 1)         (1 1 1)        32        0B        0B         -           -  GeForce GT 640          1         7                          -  mykernel(int*) [108]
1.12603s         -                    -               -         -         -         -         -           -  GeForce GT 640          -         -                     4096 B  [Unified Memory Memcpy DtoH]
1.12603s         -                    -               -         -         -         -         -           -  GeForce GT 640          -         -                     4096 B  [Unified Memory Memcpy HtoD]
1.12603s         -                    -               -         -         -         -         -           -  GeForce GT 640          -         -                          1  [Unified Memory CPU page faults]

Regs: Number of registers used per CUDA thread. This number includes registers used internally by the CUDA driver and/or tools and can be more than what the compiler shows.
SSMem: Static shared memory allocated per CUDA block.
DSMem: Dynamic shared memory allocated per CUDA block.
$