Roundtable
Neza
can make xAOD clusters from traccc clusters from pixels. Take the Athena clusters. There seems to be a dummy cell at the end, that produces a crash.
Attila: Let's look at it after the meeting.
Plan to circulate talk and poster for ATLAS-week.
Homogenious field. Ask Stephen for example. There is also in detray a unit test.
Andreas Salzburger: Add a new field-converter.
Single muons create a single track with cpu. Input were around 6-7.
Ambiguity resolution performance is good for single muons. ttbar however is realy bad.
(ttbar is same as in acts and minimises the shared hits. not parallelised yet)
Attila
Not so much coding. Started to tackle reported physics performance issues. Currently, our examples are not in the greates shape. Reconstruction with cuda on gpu and comparing to cpu doesn't match, but should be. Often it is just a +-1 error but the reporting reuslts in an error from this.
EDM and datamodel developments, because could be beneficial for the code.
Luthien
Goal: Create technical efficieny plots for single muons. We currently get errors and working on them with PF.
PF: job-options got convoluted, but created a PR to fix it. CI is still failing but should work soon.
ATLAS poster, QP in the chain.
Stephen
Followup on the performance plots: Perfromance got so bad, that we now use log-plots. Now we are slower than 2 Hz. This comes from the addition of material.
8 Hz measurments were always without material.
Andreas Salzburger: Does the number of candidates change? -> Yes
Joana: Some time ago, by adding the material the trackfinding found many more hits.
Joana: Toydetector also changes performance when switching.
Attila: The plot suggests, that we are measuring things wrong. We need a new measure for what we actually want to do.
Andreas Salzburger: we could make propagation only tests and make grid-based vs homogenious material. Then we figure out the material cost.
Algebra plugin is not yet updated.
We need to have a truth seeder. Plan to implement this week.
Andreas Salzburger: We want performance in best case seeding case.
Attila: We also want to figure out, which parts need to be improved. -> Andreas Salzburger: try all might not be a great idea.
We have 2 parameters that do the same (branching and max tracks) but they are connected. Sorting is not straight forward.
Andreas Salzburger: ATLAS runs branching factor 1. This means always take the best one at each surface. This prevents the explosion of the combinations.
Attila: If we only want the best, then we can save time on the sorting.
Andreas Salzburger: Pixels give an extrely good seed. Taking the best from there is a good approach. Might be the solution for the ITk.
(Everything on ODD with const magnetic fields)
Plans to introduce 3 new Bfields for testing:
- ATLAS-ITk field
- simple constant Bfield with interpolation
- Bfield with ATLAS binning
Andreas Salzburger
Waiting for traccc to update for the last tag. Then we finally close the gap and are up to date. Detray has been updated already.
Moved material writing to core. At the same time changed to run with root tree and nTuple backend. (No root dependencies introduced.)
Attila: latest root-version has now a fixed binary format for nTuple. This will stick forever. However, the API is not finalised yet and should come in May.
Paul: We still need to wait quite long to require it.
Probably not coming this year on our side, because we can't require everything from users. Still written as core-plugin to be picked up by ATLAS.
Material mapping files are huge with nTuple. (e.g. ITk phi-binning 10^6 tracks -> 80 GB. ITk takes 5x as long as ODD)
Noemi
New Phd students Luthien and Mikkel. Workpackage 2.4. Not bound yet to a tehnology choice to help also for PhD.
- Mikkel: Started with Benjamin on GNN site. Look onto GPU-workflow.
- Luthien: CPU-performance with Noemi and plans to go later to machine learning approaches.
Only small ACTS PR and Script for seeding performance. Plan to use vTune to check for hotspots. Script can read and feed into seeding algorithm.
Attila: vTune: We need to keep on the radar, that intel might do a non-free product in the future.
PF
PR to ACTS that changes affine transformation. Afine transformation has always to check for inverse and rotation to remove shewaring and scaling. Therefore always a bit of overhead. There is no reason to use affine in tracking. Was only a single case in ACTS where it was used. In Athena, this kills the muon part.
Run vTune on ACTS in Athena and reported 4 hotpots linked to fitter candidates before doing an operation. Removed one of the sorting. Both changes performe worse. Double sorting is very expensive but seems to be the best solution still. Now looking at other places for improvements.
Andreas Salzburger: mp-nt-tests: Where will it be reporter? We should document that.
Attila: Pure tracking jobs on amd? they sound really bad.
Andreas Salzburger: Go back to core software an report the findings and figure out there on how to proceed.
Noemi. This discussio will also important for computing TDR.
Attila will initiate the reporting and send out emails.
Doga
Large radius workflow. Reconstruction in Athena done by ACTS. Number of input seeds is similar. ACTS takes 2x time as legacy. Not clear yet, what part is the problem.
LRT: Number of input seeds to ACTS is 1/16x to legacy. Will be added in QP endnote. Discussed with Paul to make some cuts.
Noemi: Last week, trigger study people were interested in results of LRT passes and would be interested in improving together. Will forward information when received.
Analysis. Luis is on vacation, so Doga will do the presentation.
Mikkel
Started to work on the GNN. Looking at the performance only looking at the pixels and extrapolating later. Not sure yet, why there is a drop in performance for |\eta| < 1. This is the efficiency of the graph the comes out.
Alex
S&C week:
- Sustainabilty studies seem not super consise because a lot of things are neglected.
- No clear answer what language to use for writing sustainable code when user is a bad coder.
Work for University. Github classroom is
Paul
Failed to merge ACTS bump into Athena. Thrusday coming.
Thinking what parts of the core-library to break out. KF affected (?). With these refactorings we might get it down (?) And then we are good to go.
Joana
Tried to get new detray version in traccc. Failed on random numbers from a different stage.
String stuff that failed only on cuda-ci (with the same image?)
Refactored interfaces to detray navigation. Planned to write a direct navigator.
Go back to benchmarking homogenious material map. Toy detector seems to work quite well.
Andreas Salzburger: Beomki wants to talk about the mathematical changes in detray. Discussion will be for fact-finding. Related to time. Beomki could speed up propagation without time.
Andreas Stefl: Using the sympy implementation has the same effect, because it sets those comonents to 0 and negelcts them.
Joana: Throughput on gpu has a cut at 4 registers. So 5 and 6 should be identical.
Andreas Stefl
Almost finsihed up dense propagation. Pulls on position are still a bit off.
There is a problem in the material combining. This alters the material properties. We average charge number and molar density. Eg. for air+iron, it doesn't work.
