Roundtable
Alex
nothing for this group.
PF
Started to profile ACTS seeding with vtune. We lose 0.5% with sorting top/bottom.
Noemi: We need to sort them, because then we have a way to skip some.
Sorting happens in the last step before the loop.
Maybe create seeding state as a global member?
Currently, we have many ifs in constexpr.
Should we discuss to redesign this part?
Started with MR to change geometries. atm Eigen specialised on rotations.
Noemi
2 PhD students join wp2.4 EF-tracking:
- Luthien
- Mika - maybe GPU dev?
Stephen
Tried to get traccc on ITk working. Stumbled over numerous new problems. Started to fix them. Important are memory issue, phi-wrap-around, config.
Neza already used some of the changes.
We now have traccc-benchmarks per kernel. Hoepfully getting into automated benchmarking. Even error bars possible.
The large jump:
Maybe from propagation?
The commit switches to a new algebra-plugin. Very likely, that a kalman-update changes the performance. The current code prioritises convinence over performance. Looking into this with Yuki.
Alex: Isn't compiler optimising that away? -> Compiler not smart enough for that because we have too many abstractions.
Maybe also start with sympy.
Andreas Stefl: Could the jump be because we now also fit in the seeding? -> Probably only coincidally at the same height, but more related to the occupancy.
Dimitra
Working on presentation for analysis meeting.
Muon-work: Seed efficiency from micromegas. Check for every combination per truth particle. Equal to 4 hits. No combinations with efficiency 1. Since we required 4+ hits, some combinations aren't even present.
Shadow for ATLAS cavern and learned cool stuff from Paul.
Neza
traccc: Through-put tests on ITk and made efficiency plots. Now we have ACTS standalone and Athena. Now preparing slides. (On G200 pipeline).
Currently it looks like traccc and Athena clusters are the same.
ttbar pile-up doesn't work yet properly.
Andreas Salzburger: Do chunks of 3-4 particles and look into log. When there is a problem, analyse, why they are not matching up. Probably not add more infrastructure, better have it with cout for the time. 1:1-correspondance is important.
Noemi: We already have xAOD to cluster matching. Also a script exists.
Joana: Effect of constant B-field? -> We need to use a different field. We can convert ATLAS-bfield from csv-file.
Stephen: We would need to update all translation units to the new field. Maybe just a single new file needed.
Talk on Wednesday.
Andreas Salzburger
Detray is 10 versions behind. We need to get this updated, it is still working but a it of a problem. But a next great step.
Test fit with material only on sensitive surfaces works well for GNN. Afterwards having a direct navigator would also work nicely. Does only work for fitting, not for finding.
Will send plots to Neza.
Attila
Spent most time supervising tutorials.
Doga
ACTS workflow for LRTs in Athena. Looking at main pass. ACTS seems to be a lot faster, but efficiency is not great. (Not confirmed yet). Starting to look into more detailed timing, to see which functions are the heaviest. Should end QT.
Analysis: Back ground modeling.
Joana
Göttingen-HPC: Compil;ing detray, but ancient compiler and no boost. Only intel-CPUs.
Reworking old PR.
Plan on alignment discussion with Marko.
Andreas Stefl
Dense propagation: started to finish chapter for thesis. Then also did this in simpy. Now energy loss also possible there. Now we can do covariance after each step in dense material. Dense field and performance measures necessary and the chapter can be finished.
