Roundtable
Alex
Discussions with Andreas Stefl on GX2F.
University stuff (homework and job)
Booked this room for the next half year. Sometimes on Tuesday, when Monday is off.
Paul
Remove implicit constructors. More of a decision problem but we don't want them.
ttbar pu200. up to 32 we have good scaling efficiency and then we break down but still improves. Surprisingly, each added thread removes a constant. Maybe we could lend a 192 thread machine?
ITk geometry building. Pixel works. Material is randomly place. SVG-output is not working yet, needs a look into library.
Making proposal for edm4hep to store clusters in a different format with axis indices.
PF
Also performance on multitracking in Athena with ACTS. Tests on traccc-AMD-machine: Far away from linear scaling but legacy and ACTS have similar performance.
Paul: Maybe Athena malloc-library does not like changing threads. We are relying on that.
Andreas Salzburger: Organise talk in ATLAS core software and present our results. Maybe this has a larger impact than many other improvements combined. FOLLOW UP on this.
Doga
hgtt-implementation, enum-problem: The algorithm now works but the store-key makes problem because we initilise it twice. Not sure yet, who is causing this.
Started to write int-note.
tt-di-higgs work continues.
PanDA problem: jobs get stuck at running since a few days.
Paul: Maybe down-prioritised?
Noemi: Maybe not all samples available?
Andreas Salzburger
Introduce new proto-axis as followup to proto-binning. Already compiles, should be in soon.
Geometric digitisation file for the ODD. Plan to get good resolution. Strongly improve ODD.
Noemi
ACTS-integration: Opened a few JIRA-tickets to discuss some open issues.
GBTS-PR made a bit of progress. Offline discussion is necessary. Huge problem ist that often things like if strip/pixel/lrt occure. We want to have it more generic.
Andreas Salzburger: We now have also more geometries for testing. Maybe even soon the FCC one.
Roman
Here for a week. Looked through Paul's code. Plan to work together with Paul a lot and focus on navigation. Start with telescope.
Paul: desired outcome: More telescope style workruns and tests.
Dimitra
Ranked seeds in x-zero-resolution. If seed is found, removed from maximum and repeated. Also want to study seed efficiency to see how efficient each seed is.
Book-keeping in a json-file. Start with multilayer and then continue and remove the found hits.
Seed extension is also measured.
Efficiency := nConstructed/ nTotal
Run in signal reagion to check if number of events make sense (some one else did)
Neza
traccc: implosion on number of created seedsa after seeding config changed in traccc. Now trying to find a sweet spot, because too many seeds found (millions. goal is around 100k). Now making a poster for ATLAS-week.
confusion about strip-clusters.
There were 2 others working on the topic. Strip-digitisation was missing. Neza did for pixels and other do strips. It seems some users want to only press buttons instead of modifying the code.
Andreas Salzburger: This probably needs more time, until everyone has a deeper understanding of the basis.
Plan to have similar thing as Beomki presented at CHEP.
Not clear yet if high number comes from strong branching or other effects.
It could also be that parameters are wrong for the ITk.
Athena has some cuts on the middle clusters. That might change the numbers. We don't require exact matching of the numbers.
Joana
Working on thesis thing to make sure, that detray version can be exported.
Stephen
Spied on GBTS-seeding meeting. "Secret" meeting on Friday 16:00. They have already a working version for gpu which was used on run2. We have big interest in understanding, how they managed to get below 500ms.
Andreas Stefl
Merged some ACTS-Athena PRs. Finally the configuration got merged.
Navigation in ACTS was changed. In Athena, most of them had to change the error-masking. GX2F failed because got outside of well defined area.
- Andreas Salzburger: Iterative fit could do this.
- Alex: implemented a volume-check some time ago that fails when leaving the initial volume
- Athena does not allow errors.
- error-handling in algorithm instead of actor. because in actor it is too late. but also has performance impact.
Plan to do navigation fail on initialise.
Prepare talk for eIC
