How to assign GPU to a worker

There is a cluster of PCs. We have just added one GPU (CUDA 1.3).

What should I do to integrate GPU into cluster? Or how can I communicate with GPU from "cliens" PCs?

[Daniel Shub]

This is a cool question. I am not sure if you can add a GPU as a worker to a cluster. I think it will probably require adding a work CPU that has access to the GPU. You might have to add the entire machine to prevent the GPU from being a shared resource.

Answer 1

[Jason Ross]

As you don't specify the cluster you are running, I can't be all that exact, but you need to ensure that that your GPU-related work ends up on that machine. Then, when GPU code runs, it will use the GPU.

It can also matter what operating system you are running on the systems in the cluster, as the GPU needs to be able to run in "service" mode to be able to be used without someone logged into the desktop. The actual configuration varies by OS, scheduler and GPU driver, so I can't be specific here, either.

Answer 2

[Alexey]

Let it be 2 PCs with identical characteristics. OS: windows 2008 x64 Each PC has GPU(Cuda 1.3)

I would like to create job manager and parallel tasks between GPUs.

There is a chance to buy Tesla GPUs (which should support parallel computing) and put them on both PCs.

Is it possible somehow to create workers with assigned gpu?