Velo-SciFi track reconstruction based on SciFi seeding and Matching algorithms

added RTA label

From the first round of nightly tests this is the like-for-like comparison on the v6 SciFi dataset. (Actually given the low-momentum efficiencies the real like-for-like comparison would be with the forward without cuts, but let's leave that aside for now) Looks very positive. The reduction in performance on CPU is maybe something to think about but not to me a huge deal.

@dcampora @dovombru @decianm given the unexpectedly positive results I'd appreciate an expedited review here with a focus on figuring out whether there is any major mistake/bug which spoil the story and would prevent this being used. We need to be in a position to report on this in LHCb week.

Seeding + Track matching

NVIDIA GeForce RTX 3090    │████████████████████████████████████████████████   244.16 kHz
NVIDIA RTX A6000           │███████████████████████████████████████████████    239.77 kHz
NVIDIA RTX A5000           │██████████████████████████████████████             190.43 kHz
NVIDIA GeForce RTX 2080 Ti │████████████████████████████████                   160.37 kHz
AMD EPYC 7502 32-Core      │██                                                 14.79 kHz
Intel Xeon E5-2630 v4      │▌                                                  3.25 kHz
                           ┼────┴────┼────┴────┼────┴────┼────┴────┼────┴────┼
                           0         50       100       150       200       250

Velo + VeloUT + Forward

NVIDIA GeForce RTX 3090    │████████████████████████████████████████████       223.53 kHz (1.00x)
NVIDIA RTX A6000           │███████████████████████████████████████████        215.68 kHz (1.00x)
NVIDIA RTX A5000           │███████████████████████████████████                178.03 kHz (1.00x)
NVIDIA GeForce RTX 2080 Ti │████████████████████████████                       143.02 kHz (1.00x)
MI100                      │██████████████████                                 94.57 kHz (1.02x)
AMD EPYC 7502 32-Core      │███                                                16.34 kHz (1.00x)
                           ┼────┴────┼────┴────┼────┴────┼────┴────┼────┴────┼
                           0         50       100       150       200       250

@maxime would it be possible to test whether this has any negative impact on the track-cluster matching, electron reconstruction, and brems recovery enabled in !565 (merged) ?

The automatic output rate tests on minimum bias samples confirm the increase in TrackMVA rate shown in the WG presentation. As @sstahl pointed out this is unexpected and could indicate a bug in one of the quantities this line cuts on -- PT, track chi2, or IPCHI2. Hints are that no big increase is seen for the two track line and TrackMVA does not seem to increase the output rate on a Bs2PhiPhi sample.

While investigating this I would suggest to run some line independence tests turning off everything except the TrackMVA line (or the TwoTrackMVA line, or ...), because we anyway need to run those before datataking to make sure we haven't introduced any framework bugs over the past months so we might as well try to save some time by dealing with that as a side-effect of figuring this out.

added 1 commit

88a482aa - fixed copyright issue

Velo-SciFi track reconstruction based on SciFi seeding and Matching algorithms

Merge request reports

Merged by Rosen Matev 2 years ago (Aug 31, 2022 7:37am UTC) 2 years ago

Activity