Alice Weekly Meeting: Software for Hardware Accelerators / PDP-SRC - MINUTES ONLY

Wednesday 5 Feb 2025, 10:00 → 11:30 Europe/Zurich

1 Discussion

Speakers: David Rohr (CERN), Giulio Eulisse (CERN)

2 Following up JIRA tickets

Speaker: Ernst Hellbar (CERN)

3 TPC ML Clustering

Speaker: Christian Sonnabend (CERN, Heidelberg University (DE))

GPUmeeting_05022025.pdf

4 ITS Tracking

Speaker: Matteo Concas (CERN)

5 TPC Track Model Decoding on GPU

Speaker: Gabriele Cimador (Universita e INFN Trieste (IT))

GPU param optimisation

Updates:

• Performed several measurements for independent kernels (MergerTrackFit, MergerFollowLoopers, MergerSliceRefit, MMergerCollect, CompressionKernels_step0attached, CompressionKernels_step1unattached)
• Created set of best parameters per beamtype (pp, PbPb) and per IR (100kHz, 2MHz for pp and 5kHz, 50kHz for PbPb)
  • How to choose best configuration:
    1. compute conf_mean_time - default_conf_mean_time
    2. propagate error (std dev) of the difference and compute 95% confidence interval
    3. if 0 is in the interval, can't tell with confidence if current configuration is better than the default
    4. if one or more CIs have upperbound < 0, choose the one with smaller mean (i.e. the best)
• Started to investigate on concurrent kernels
  • TrackletConstructor_singleSlice params, as it is the overall longest kernel with other concurrent kernels
  • Grid search on its grid and block size
  • Measured both mean kernel duration and overall TrackingSlice duration (changing params influences whole phase duration)

 To do:

• Plug in the optimal set of parameters per (beamtype,IR) and check if there is a noticeable improvement in TPC processing when input TF is (beamtype,IR)
• Continue to investigate on phases with concurrent kernels (more difficult than single stream kernels)

Problems

EPNs are currently down.

6 Efficient Data Structures

Speaker: Dr Oliver Gregor Rietmann (CERN)