Joint meeting for GPU validation

513/R-068 (CERN)



Show room on map
Joint meeting for GPU validation
Zoom Meeting ID
Zenny Jovi Joestar Wettersten
Useful links
Join via phone
Zoom URL

Wed 24.05.2023 PhD meeting
Present: Zenny Wettersten, Stefan Roiser, Stephan Hageboeck, Andrea Valassi, Robert Schoefbeck, Jin Choi

# Status

ZW: Jin which tests have you run?
JC: tested what ZW gave for reweighting, but would need other processes
Also had a look at SR's gridpacks. But they are very different from CMS centrally produced gridpacks.
SR: discussed with Sapta, SR is now integrating some latest developments and will produce new gridpacks

SR: which cuda and gcc?
JC: CUDA11.5 and gcc11.something

SR: new gridpacks will be in /afs/
(security team did not like the previous publicly available AFS area)

SR: note, we might need some extra performance optimizations

RS: is it standarda gridpacks? different from CMS gridpacks?
SR: whatever comes from madgraph, but in addition with gpus

JC: is there a plan to port this to the CMS production?
RS: yes sure, but we need to test it first
SR: looking for some functional testing for the moment

# Next steps

ZW: what are the next things you would like to see?
RS: Sapta could not be here... anyway timing improvements would be a very important step
This could be used for our HL-LHC CPU consumption estimates
Sapta suggested DrellYan with many jets, I suggest 4top with many jets

RS: how applicable is this to pdf weighting?
AV: not sure, we mainly speed up the matrix element evaluation
RS: we store many pdf weights, not sure if we need ME reevaluations, probably not?
maybe through the renormalization scales it would require ME reevaluation?

RS: how about EFT?
ZW: we have some issues with code generation for EFT for the moment

RS: you could launch 1000 times the same reweighting 
(NB: this is launch for reweighting "launch --rwgt....")

ZW shows some slides from CHEP
Slide 11 there is a normalization at 10^2 events (AV: not very clear why)

JC: Sapta suggested Drell Yan, I am not sure if the reweighting is relevant

AV: we are focusing on a full unweighted event generation now and SM, 
then EFT reweighting we are doing at lower priority, is this ok?
RS: yes perfectly fine, with SM and with full workflow

AV: Jin when you say CMS gridpacks you mean unweighted egenration or also reweighting?
JC: both, we also have gridpacks for reweighting
RS: look in particular at

Discussion on Drell Yan
Without extra jets it would be too simple, but with more jets it would be useful
SR: Sapta gave us
RS: yes this has 4 jets

[SR and RS have to leave]

AV: could you maybe send us some instructions for a reproducible way to run
JC: sure, we have a twiki but it may be protected, I will look again and will follow up later


There are minutes attached to this event. Show them.
The agenda of this meeting is empty