Commissioning at P2

• After bug fix: Successful runs at P2 with all algorithmic approaches. Some stats from the logs (tRateLimit = 240, 500 kHz pp replay)
  
  
  • Run without backpressure and CF regession, threshold 0.03:
    • MI50 (nEPNs = 71): 2.652 +- 0.001 TF/s/EPN
    • MI100 (nEPNs = 45): 2.997 +- 0.003 TF/s/EPN
      
      
  • Run with backpressure and NN regression, threshold 0.03:
    • MI50 (nEPNs = 72): 2.119 +- 0.001 TF/s/EPN
    • MI100 (nEPNs = 44): 2.440 +- 0.002 TF/s/EPN
  • Run without backpressure and NN regression, threshold 0.05:
    • MI50 (nEPNs = 78): 2.175 +- 0.002 TF/s/EPN
    • MI100 (nEPNs = 50): 2.541 +- 0.001 TF/s/EPN

• All runs succesful, no memory access faults, no (clusterizer induced) crashes. In the NN regression: Few runs still showed the PAD IROC-OROC out-of-bounds error (O(20) messages total). To be investigated, although not reproducible in the FST even at high rate, so it might even be legit.

Physics

• Discovered a problem in analysis on real data -> Fixed and rerunning now
• Analysis will follow by the end of the week (if all goes according to plan for job submission)
• ITS-TPC matching efficiency investigation: Pending. Real data & PID analysis has priority this week.
• Sanity check on low rate pp data: Anchored to reference run, minimum bias, no SC distortions (because they are basically negligible with this pp data). Metrics improve or stay unchanged using the networks trained on PbPb 50kHz. Training on SC or no SC data has no visible difference on performance. Most notable difference: Efficiency vs pT for secondaries (first plot) and chi2 for good tracks (second plot). Fake rate stays approximately the same.

Framework

• One more fix needed in O2/dev to run with MC data. https://github.com/AliceO2Group/AliceO2/pull/14677

Otherwise mainly busy writing, correcting, rerunning