Roundtable
Alex
Athena GX2F
Thesis
Plan to do the GX2F refitt of KF tracks.
Doga
EB (editorial board) meeting on Thursday went well. Work was on tt-di-higgs-paper. This is also relevant for the thesis.
Going back to hddt work.
Luthien
Checked efficency proposed by Carlo's fix.
Plan to learn how to make the material mapping. Noemi will do the other option.
Dimitra
Validation of tracking geometry in ATHENA. Almost all tests in place. Tested with ATLAS extrapolator. Many hits are off comparing to simulated hits. There was a magic factor of 1000x, because the magnetic field was too small. Removing it makes it worse. Still trying to understand the meaning.
Outliers with 30-40mm. Compatibility was expected. Maybe stepsize would be next step to investigate. Pathlength is many meters.
Andreas Stefl: Maybe fix step-size to 10mm
Navigation discussion: Planning on offline discussion after creating a PR. Index as broken and can get into ATHENA with next update.
Noemi
Working on material mapping with Luthien.
NGT side: Dirac-week and working on slides/managing to get all the slides from all people. Meeting will be on Tuesday.
Reminder: prepare slides earlier!
Andreas Salzburger
Grids:
Problem: everything in the code got templated. For running, it was no issue, but I/O, we had to go through all steps to get the correct type. With axis, this was already solved. Code to write out on the grid based material stuff hat a lot of unrolling.
Solution by Paul: No virtual calls on the axis, instead on the value type. Now with type erasure. Speed in this part of the code is of no concern, maintainability is the goal.
Neza
Profiling traccc in ATHENA. ACTS was a bit slow. GPU-roundtable discussion on how to do the throughput measurements. Booked some time for testing.
Discussion on what to include in the GPU-document. It is not clear yet, what we want to tell with it. How to include GNN-development? Maybe just a single paragraph? Because another report would be too excessive. The report template is already 45 pages. Continue discussion offline.
Attila: GPU and FPG measurements should come to an agreement. CPU is different and hard to mess up.
Andreas Salzburger: With the testbed, we want to show that what-ever solution gets selected is a stable and feasable method.
Presentation will be 25 minutes. Plan to circulate slides next week.
PF
Plots: Technical efficiency: Difference between analog and digital clustering with different chi2 cuts. Still suffering outside -3 to +3 eta.
Noemi: Volume cuts help to rely on position of the seed instead of the type. This should be helpfull for high eta.
Fake rate similar to legacy already. We can also see increase of resolution in the barrel.
Bug in TOD to charge=0 when overflowing. IPL also has an overflow, which also would effect run3.
Attila
HSF cpp tutorial.
offline IO patterns. Hoping to get back to EDM in traccc.
Mikkel
GNNs finally in ACTS. Hopefully this works out this week.
Joana
Scaling plots: Scaling looks great, until first CPU is saturated. Scaling breaks between 32-64 threads. Either scheduling system doesn't give second CPU or NUMA (non-unified memory access) makes problems.
Attila: NUMA should be only relevant, when CPU-threads need to communicate with each other. More likely, that scheduling is wrong.
Algebra plugin benchmarks. VC-stuff was slow in detray, because crossproduct was not vectorised yet.
Debug navigation and validation tools.
Andreas Stefl
Started to integrate dense material propagation into ACTS. Might be useful for NA60+.
Writing and correcting thesis.
