Roundtable
Luthien
Anton
Noemi
Workshop registration closed. Please hand in abstracts if you want to present anything.
Paul
TBB and block-range for parallel form. Block is bad for writing events in order because we need to buffer up to 50 events to write them out in the correct order.
Attila: ATLAS does not do it.
Paul: I already do it and we have reproducability
Attila: Why not single thread for reproducability? -> Because we cannot test all desired cases.
Solution: keep cue of events. Terrible with TBB. There is no good reason for TBB, opened PR to replace it. Random stalls in threads stop other threads, if they are too far away.
Thread-assignment is not deterministic, but this is fine. Pythia is also not deterministic.
Not sure, if there is a Pythia solution in ATLAS, could also be that we are using it wrong.
Also for HepMC event numbers should be correct.
CI-updates: Spack-install cach hits are guaranteed and work. Spack-blobs are even docker-images with build-artifacts.
Caveat that images are x86 so we need to rebuild for ARM. But then we are close to "just use it"
Attila
NextGen Hackathon last week. Useful to some people but not sure, if it was the most efficient for all.
Pulled in wp2+3 people into wp1 to get additional input. ML attracted numerous people from ATLAS and CMS. Ended up to be a tutorial for ML-tools.
EDM-workpackage: CMS seems to have done useful stuff during hackathon.
Event generation GPU: Only 2 pnew eople showed up but were lead there due to misunderstanding of the pacakge.
Co-routines: 2 afternoons of meaningful discussions on scheduling on gaudi and similar.
- Learned more about co-routines and maybe they could be a reasonable user-interface for GPU-tools in ACTS.
- Nice to have a function with multiple returns and checks if it is ok or not. Interesting for asynchronous calls and to offload single things on threads.
- Interesting because it could be self-contained in ACTS and could communicate with gaudi and similar. Otherwise we would need to agree on many conventions.
- Paul: why not TBB?
Alex
Mainly studying. Looking for test-subjects to present some proofs.
Noemi: Stefano Camarda could be of good help.
Benjamin
Midcycle-presentation tomorrow: Interesting results with IDPPM but not enough time for this presentation.
Also improve efficincy for ttbar-events.
Modular graph library from Toulouse. Training is also done by other community. Just end-implementations is only Benjamin.
Joana
Benchmarking. Testing forward-backward navigation without KF. Wobbling seems to come from measurement variance. Setting the measurement variation to zero removes this effect and make sthe KF more stable.
New segfault in toydetector seems to come from the stepper with a track-state containing onnly NaN. Crash happens on the CPU.
Andreas Stefl
Thesis work and corrections.
Discussion on resolution degradance in high eta. Covariance inflation might not be accurate there. Also depends on momenta. PF investigating it this week.
