Roundtable
Alex
Continue work on Athena GX2F. It is getting clearer.
Paul
Gen3
- ID-asignment working
- svg output working
- new type to interact with the grid. Look-up-performance was an issue.
- Athena branch updated for next ACTS version, to also work with ID-asignment.
- Probably, still a lot to deduplicate. Especially in containers.
apogee CI working after 2 weeks. Jet-classification in Athena made problems.
vtune should be easier to use in the new setup
ODD was broken but works now. Problems with LCG107-root-version.
Attila: There were also other outage problems. Maybe we should report the problems.
Doga
Working on analysis. Today, final QP presentation. Just a few points missing.
Plans: Working for EP meeting next week
Luthien
Chi2 of the holes in ACTS. Examine tracking efficiency for new material distribution prepared by Noemi.
PF: What happens to measurements after calibration on surface? -> Position can only corrected with neural network. So, only covariance is changed.
PF: What happens to covariance after calibrator? -> LG probably checked during implementation. But it is unclear, if calibration is even used in ITk. Probably the same solution as in ACTS and using the same maps. Measurement can't be dropped after calibration. -> PF will recheck that.
Andreas Salzburger: ITk calibration: we use more information to improve clusters. Clusters are not physically split but made shared. It should improve the data-taking.
Mikkal
Plan: Use GNN only on Pixel and extrapolate on Strips. Size difference should be significant, since it scales with npoints^2
PF
A lot is IO-bound. The IO seems to be untouchable. Maybe there is a way to ignore that constants.
The lookup in the pixels seems not linear. Might come from different size of pixels.
Charge calibration structure is convoluted.
Paul: depending on cluster stores cell (heap to heap allocation), we could change data structure. Maybe a vector with allocators? -> Attila: This already exists.
In the last few weeks, we improved the performance by almost 10%.
Neza
Helped student build seeding efficiency and it is getting close to 1.
Attila: seeding efficiency had a lot of dips, why not anymore? -> Maybe it was a relic from finding.
Attila: In trackfinding, a further track killer could be introduced. eg for same measurement twice. -> This should be an open PR from Stephen. There might be a larger underlying problem.
Dimitra
Seeding for micromegas in Athena merged.
Seed ranking with resolution. Rank with json file and build seeds in that order. Used hits wouldn't be used for other seeds.
ACTS Gen3: started to trapezoid shaped portal
Paul: connectivity is probably not necessary. We should not do it until we actually need it.
Andreas Salzburger
Proto axis is merged now. Grid surface material will be changed into new structure.
Started PR on disentangling python core/examples/plugins
Core: some can be used/built without using the examples at all.
It is not fully clear, in which direction we want to go for everything. Open for discussion.
Attila
Space point migration concluded with not a lot of fan fair. The performance was not fully reliable, since we also made a few changes at other places. But now it should be easier to move data back and forward.
Memory managment in traccc: Currently we make always copy of measurements in memory. With the duplication in the tracks, this allocates a lot of memory.
Andreas Stefl: In ACTS the measurements are copied, that after calibration, they are unique to each track.
All steps of the fit create an object built around the last measurement of the track adding additional information. Instead of copying, we might put there in also the measurement pointers and calibration information.
Discussions with GBTS: They claim Athena should have something functional. Whole seed finding runs in a few ms. But filling the input takes already 10 ms which is a huge bottleneck. Looking forward on how this resolves.
Joana
Smoother was segfaulting on toydetector. Non-invertible matrices in KF. not from algebra plugin. It seems to come from the covariance inflation. Not sure yet, if the physics results are true still.
Research on KF and stability.
Attila: Maybe we need to keep double precission.
Double precission in inversion didnt help directly.
We then also come into performance considerations.
Alex: Does the skipping of the inflation make the effect only less likely and we might see it with different seeds? -> This is possible. The problem might be only easier visible with the inflation.
Göttingen-HPC: Problems with node-reservation causing trouble in benchmarking. Exklusive flag seems to make jobs wait forever. Plans on Communicating with them.
Andreas Stefl
2 more PRs for ACTS for this version. Should be added in soon to avoid breaking problems.
Thesis writing.
