The CUDA change the CPU affinity of my application.

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi guys,

I have a question which puzzles me for a long time:

I am using LabVIEW to call into my DLLs containing some CUDA functions, but after the function completes, I found the LabVIEW is bundled to the CPU 0, while the CPU 1 got unchecked… Please see the attachment. Both the CPU 0 & 1 will not get unchecked if the DLL does not contain the CUDA code.

This is really an unwanted behaviour for me since I want the GPU and CPU to work together, setting LabVIEW bundle to CPU 0 just make the CPU 1 idle and this is unacceptable for my application. Although I can use the Windows API ‘SetProcessAffinityMask’ to put my LabVIEW back onto two cores, but I consider this only a workaround… Does anyone have any idea about why the CPU affinity can be automatically modified by CUDA and how to prevent this? Thanks.

2

The profiler does this for accurate timing.

3

Does this mean that I will NOT see the phenomenon once I close the profiler? Thanks.

And I will try it again.

4

Correct.

5

Hi tmurray,

I tried running the DLL without opening the ‘CUDA Visual Profiler’, but that problem still persists. Am I missing any operation here? Does the profiler you pointed out is the ‘CUDA Visual Profiler’?

Could you show me a way to ‘turn off’ the profiler?

Thanks again.

6

Some of the cutil timers set process affinity as well. Are you using any of those?

–Cliff

7

Hi Cliff,

In my simplest code, I didn’t use any cutil timer, I have attached my test code below:

[codebox]void test()

{

float *data;

cudaMalloc((void**)&data, sizeof(float)*10);

cudaFree(data);

}[/codebox]

Once I finish running the code above, the affinity of LabVIEW is modified…

8

I strongly suspect the affinity is set when the CUDA runtime engine is loaded, see the code snippet below:

[codebox]void test()

{

float *data;

PrintOutAffinity("Begin the program");

cudaMalloc((void**)&data, sizeof(float)*10);



PrintOutAffinity("After the malloc");

SetAffinity(3);

PrintOutAffinity("After reset the affinity");



cudaFree(data);

PrintOutAffinity("After the free");

cudaThreadExit();

}[/codebox]

The function ‘PrintOutAffinity’ prints out the current affinity of the process, and the function ‘SetAffinity’ just set the process affinity to 3. The execution result is shown below:

[codebox]Begin the program 3

After the malloc 1

After reset the affinity 3

After the free 3[/codebox]

It’s clearly seen above that the affinity is only set once at the first CUDA function call, if I set the affinity back to 3, and it will never be modified again. So I think the affinity is only modified once at the same time with the CUDA runtime engine load.

Sound reasonable?

9

Do you have the CUDA_PROFILE environment variable set to 1?

10

Hi tmurray,

You are the man!!! You just help solve a problem which bothers me for a month!!! Great thanks.