Cuda FFT Plan Creation Slow Over Network

I am encountering a strange problem related to the performance of Cuda FFT plan creation here at my workplace. Since last week, our testing department has been reporting that the routines in our software that depend on FFT computation have slowed down significantly. We diagnosed the problem down to abnormally long times for the FFT plan creation.

What is strange is that when we disable the LAN of the PCs, the FFT plan creation becomes fast again.

So my question is: What is (if any) the dependence of CUDA FFT plan generation on network connectivity. If not, what could we possibly be doing wrong here?

We have tried using the sample CuFFT application “simpleCUFFT” and the observations are exactly the same - slow with network on, fast with network disabled.

Thanks

CUFFT plan creation should not depend at all on the network. However it will depend on CPU throughput. So if your network is slowing your system down, that could be an issue.

Regarding this statement:

“Since last week, our testing department has been reporting that the routines in our software that depend on FFT computation have slowed down significantly.”

Presumably something changed in your setup last week. You should probably identify the specific change that causes this.

Thanks for the reply txbob.

Something has probably changed in our network and we are working on identifying that. What still perplexes me is that our application uses several kernel calls and yet the slowdown appears to be only on the routines that use cuFFT. What do you think is causing this? Or do you think we have slowed down in general and have not noticed that due to the small execution times of those kernels.

Thanks