Use Putty to access GPU on Lan How to use it?

Hi all,

I would like to access GPU which is installed on other computer(fedora10 is installed on it).I heard with the help of putty we can access GPU on Lan and run our CUDA programs easily.But I have following question about Putty:-

    There is any compatibility issue of putty with CUDA?

    Is thereany limit on number of users to access GPU using this particular software(putty)?

    I would like to know about your experiences putty with CUDA.

Thanks

With any ssh based remote access (putty is an ssh client) and CUDA, you aren’t “accessing” the GPU on the LAN, you are just running a remote terminal of the machine hosting the GPU. If you don’t need to use the GPU for on screen rendering, it will work fine with CUDA. All the same caveats and restrictions that exist if you were sitting in front of the machine rather than over a remote terminal session.

There exists an open source project called rCUDA that forwards CUDA API calls via a TCP socket. It may have some restrictions regarding OpenGL/DirectX interoperability, and of course it adds evil cudaMemCpy() latencies. Otherwise it should be fine to use.

Not quite PuTTY, but it allows for a true “remote access”.

URL:
http://www.hpca.uji.es/?q=node/36

Christian

No, but other users should be your first guess in a shared GPU environment for questions like “Why is my program running slowly?”, “Why did my program run out of memory?”, etc.

It is also worth noting that context switches between GPU users on one card have overhead, so a 100% busy GPU in use by 4 users will have lower throughput than a 100% busy GPU in use by only one user.

Our team have been using CUDA over SSH for long… Even on windows, we installed Cygwin and SSH Server and everything worked cool…

The only thing is : When you want to measure peformance or time sensitivity (as others point out), make sure you are the only one who is using the system. Thats all. If you are a small team, this should not be a problem.

hi I would just like to ask how to actually run the SDK (or any of my programs) using the remote GPU? what do i need to do to integrate them? Thanks!