NPPS Leadership Team

[Brett Viren <bv AT bnl.gov>]

"Laycock, Paul via Phys-npps-mgmt-l" <phys-npps-mgmt-l AT lists.bnl.gov>
writes:

> LDRD Type A - DUNE part 2 could be a good match but reading this now I
> don’t know how much overlap there is with the HEP-CCE work.. Brett ?

The overlap would be one of our making or breaking.  We can keep the
work away from HEP-CCE/PPS.  Or, with not much nudge, might be taken up
by HEP-CCE/PPS.


Maybe too long to read, but more info on what that means:

The entry point here is that HEP-CCE/PPS folks are recently interested
in "GPU batch processing optimization" (note: not panda/condor's
definition of "batch").

This "batch" is mostly about sending ever larger contiguous chunks of
data to ever more sophisticated GPU algorithms in order to fight
throughput losses due to the latency involved in doing the same overall
algorithm but with smaller chunks sent more frequently between CPU and
GPU and back.  There is one way to use WC sim to do "A/B" tests of the
two approaches and that is being considered in PPS.  Frankly I don't
expect any super interesting results.  But, the work would likely have a
side effect of getting more WC algs working on GPU which would be
welcome.

But then, "GPU batch optimization" can also be interpreted as how to not
stall-out the GPU for lack of input.  This interpretation also came up
in the PPS discussion and it leads right to Wire-Cell's "GPU service"
idea.  So far I've tried to hold my tongue in the PPS discussions to
suggest WC GPU service as a topic of work in case EDG, NPPS or others
"closer to home" could work on it.

There is some special concern here.  FNAL has already done some work
using some proprietary nVidia GPU-as-a-service package.  OTOH, WC's
solution uses open protocols.  WC's started first but has a tiny
fraction of the effort that FNAL's put up.  So, they already have a note
out and we have prototype software.  Then, FNAL has representatives on
HEP-CCE/PPS.  So, I worry that pushing the "GPU service" idea in PPS
turns into "apply FNAL's idea".  I'm not at all interested in that.


In any case, I think DUNE and other HEP experiments need *something* in
this area to efficiently use CPUs+GPUs in a production setting.  I think
an LDRD case could be made.  Maybe Type B.  But, I do have doubts if
Physics would want to support one.  Getting a feel from the department
on their will in this direction would be a very useful outcome of this
strategy request.

-Brett.

Attachment: signature.asc
Description: PGP signature