HEP-SCORE deployment TF meeting #30

Wednesday 15 Jun 2022, 16:00 → 17:00 Europe/Zurich

Zoomland

Domenico Giordano (CERN), Randall Sobie (University of Victoria (CA)), Randy Sobie (University of Victoria (CA))

2022-06-15-HSTF-minutes.pdf

draft minutes

1 Welcome, note-taking, notes from previous meeting

minutes of previous meeting [PDF]

2 Status of benchmarking campaigns

Speaker: Gonzalo Menendez Borge (CERN)

2022-06-15_TF_Measurement_Collection_Update@TF.pdf

2022-06-15_TF_Measurement_Collection_Update@TF.pptx

CPUs_missing_entries-2022-06-15.png

Entries per CPU-data-2022-06-15 11_17_48.csv

3 Progress on workloads

Speaker: Domenico Giordano (CERN)

perfMon dumpt HT ON(left) HT OFF (right).png

1.  Alice gen-sim-reco
   1. new code released by the Alice experts (thanks to Sandro Wenzel and Stefano P.)
      1. latest improvements of the Alice reco are included
      2. breakdown of time report in sim, digi, reco parts, useful to decouple the different contributions in the json report
   2. still issues to fix to get a standalone container
      1. the new code includes queries to external services to retrieve meta-data. 
      2. A fix will be released to avoid that (otherwise the standalone container cannot pass the no-network test)
   3. Feedback about zombie processes (see previous report)
      1. Asked by email for feedback if other sites have seen these zombie processes
      2. Received answer from PIC (thanks to Carles Acosta)
         1. script Dv1 (first version of Alice code) had these processes, not seen for script Dv2
             
2.  Atlas reco MT on real data
   1. during tests with HT enable on a server with limited memory (a not ideal configuration) it has been noticed that, although the duration of the process was x4 longer of the HT off case, the scores were equal.
   2. further investigation has highlighted that the CPU time was used to measure the score, instead of walltime (see comparison HT ON/OFF)
   3. Ongoing Actions
      1. fix the code for a new relase
      2. contacted all WL experts to confirm if walltime or cputime is used in their score definition
      3. checking the effect on the "Matrix" measurements. Should be small, if enough memory is available in the server. In that case the CPU efficiency is of 97%

4 Survey on HEPscore expectations

Speakers: Randall Sobie (University of Victoria (CA)), Randy Sobie (University of Victoria (CA))

Sobie-TF-Survey.pdf

5 Any other business

Calendar

• HEPscore Workshop (19-20 Sept 2022) https://indico.cern.ch/event/1170924/
• Next meeting and Summer break
  • proposal: move the August only meeting from 17th to 24th, to facilitate attendance