Cluster finder

• Successful Pb--Pb runs last friday 🥳
• Some preliminary results from cpass0:

• 16% CTF size reduction

NN improves separation power by ~10%!

• DCA distributions maintained
• Shared cluster map very similar
• NCl/track shows the expected reduction: Approx. 2-3 clusters for long tracks

GPU benchmarking

• PR: https://github.com/AliceO2Group/AliceO2/pull/15514
• Benchmark single GPU performance across servers (testing on epn000 and dev00 server)

• Multithreading CPU processes for ITS reduces downtime between gpu-reco processes enough to make measurement
• Top: Default reco, Bottom: NN full reco. One TF of LHC24ar. TFLOOP=100
• Allowed downtime between gpu processing steps: 50ms

All detectors, default settings

• epn000

• pdp-dev00 (RTX5080), single GPU:
  • [1] Downtimes too large for default reco (80-130 ms even when setting CPU process threads to 24) -> Best guess-timate
    • Wall-time mean including allowed gaps: 0.390235 s
      Individual duration sample mean: 0.364311 s
      Individual duration sample sigma: 0.00405494 s
  • [2]
    • Wall-time mean including allowed gaps: 0.607306 s
      Individual duration sample mean: 0.587373 s
      Individual duration sample sigma: 0.0101604 s

Summary:

• Processing time increase using MI50 NN / default reco: ~1.5
• Processing time increase using RTX5080 NN / default reco (to be taken with a grain of salt): ~1.5

-> No improvement (so far) for NN using RTX5080 over MI50 (concerning hardware acceleration)

Measurement with "WORKFLOW_DETECTORS": "TPC,CTP"

• epn000

• pdp-dev00 (RTX5080):

Summary:

• Seems much more precise and with less losses than using all detectors. Can load GPU much better (way less losses of TFs due to exclusions...)