CERN Accelerating science

GPU support in SWAN

Hi all,

First of all, I am glad to start using SWAN in my projects at CERN and would like to thank all the developers for the efforts made so far!

Looking at some of the previous posts on this forum, there was a reference about GPU support in SWAN a few months ago. Is there an approximate launch date for this yet? I am planning a year long project which might benefit from GPUs, so it would be help us to have even a vague idea on when this might happen.

Hi Christian,

Thank you for your kind words!

We are very close to having a prototype SWAN server with GPU support. This means you can start a session with a software stack that includes Cuda and the GPU-compiled version of tensorflow. When you do that, a GPU is made visible to your user session and, if you use tensorflow from your notebook, it will offload computations to the GPU.

We can let you know when the prototype is ready (should happen in the next one or two weeks) so that you try it out and give us feedback. On the other hand, we would like to know about your use case: what group you belong to, what you want to use the GPU from SWAN for, what libraries you need, etc…

Cheers,
Enric

Hello Enric,

Thanks for the good news!

Could you build pycuda and pyopencl packages?

Which type of GPU would be available?

Cheers,
Riccardo

Hi Riccardo,

So far we got a VM from IT for testing purposes with the following card:

NVIDIA Corporation GV100GL [Tesla V100 PCIe 32GB]

What is your use case for GPUs? Could you elaborate a bit on who would use them from SWAN and how?

Cheers,
Enric

Hi Enric,

In my case, as part of the B-train renovation (TE-MSC-MM) we are planning a neural network model using tensorflow, which will have to be trained with a large amount of data. It is still very early stages for us but we are willing to participate and try out any prototypes.

Christian

Hi Christian,

Ok, I will then keep you posted about the prototype.

Cheers,
Enric

In my group there are several particle tracking codes that are being developed for GPUs, for which the V100 are particularly strong due to their memory and double precision performance. We use primarily Python to set up simulations and use pyopencl/pycuda (recently cupy) both for development and production (as well as compiled code). The effort GPU has just started, therefore the usage would be limited at the beginning, but I would find useful to be able to run these code for demonstration and development in SWAN.

Hi Riccardo,

Thank you for the explanation. We will keep you posted as well about the prototype.

As for the inclusion of pyopencl and pycuda in the software stack, I would recommend that you open a ticket to the librarians:

https://sft.its.cern.ch/jira/projects/SPI

Please describe the packages that you need, why you need them and potential number of users. Also please specify that you would like those packages to be included in the Cuda10 stack that SWAN is using.