NPPS Leadership Team

[Alexei Klimentov <alexei.klimentov AT cern.ch>]

BTW, why not to expand it to Google (not only HPC)

We propose a collaboration between CSI's HPC group and the Physics Department's NPPS group to combine their expertise towards developing, deploying and applying distributed machine learning services. CSI's HPC expertise combined with NPPS's cloud expertise and experience building ML services using PanDA and iDDS and working with physicists in identifying and supporting physics analysis use cases will enable us to effectively leverage large scale HPC and commercial clouds resources for ML to a much greater extent than either CSI or the NPPS team can on their own.

On May 19, 2021, at 5:47 PM, Brett Viren via Phys-npps-mgmt-l <phys-npps-mgmt-l AT lists.bnl.gov> wrote:

Hi Torre,

Torre Wenaus writes:

I believe she's asked Brett to participate too.

Yes, I've just read it.

It looks good but I think we need to sharpen the argument a bit.

For example, billing it as an "AI/ML service" puts it in competition
with established SaaS (AIMLaaS?) eg such as provided by Google.  We'll
need to indicate why our idea is novel.

To me, what is novel is that we need GPUs and other special hardware in
the otherwise "historical" HTC environment.  We need GPUs not just to
train AI/ML and run AI/ML inference but also to accelerate "mundane"
algorithms (eg FFTs in DUNE's case).  Training can be done in HPC but
these latter two are applied in HTC.

More generally speaking, I think we are working to balance the "ratio of
ratios" (CPU/GPU required by software vs provided by hardware).

We should invent a new buzzword for this marriage of HTC/HPC ("HT/PC"?)


-Brett.
_______________________________________________
Phys-npps-mgmt-l mailing list
Phys-npps-mgmt-l AT lists.bnl.gov
https://lists.bnl.gov/mailman/listinfo/phys-npps-mgmt-l