HEPiX Benchmarking Working Group

Friday 6 Dec 2019, 13:55 → 15:00 Europe/Zurich

31/S-023 (CERN)

Manfred Alef (Karlsruhe Institute of Technology (KIT)), Domenico Giordano (CERN), Michele Michelotto

14:00 → 14:20 HEP WL validations

Speakers: Domenico Giordano (CERN), Manfred Alef (Karlsruhe Institute of Technology (KIT)), Michele Michelotto

BmkWG-meeting-6-12-2019.pdf

14:20 → 14:40 HEP WL validations (II)

Speaker: Michele Michelotto

Spread-all-workload.pdf

Spread-all-workload.pptx

14:40 → 14:45 HEP WL validations (III)

Speaker: Manfred Alef (Karlsruhe Institute of Technology (KIT))

• Still observing spread of up to more than 15% in Atlas gen workload (1 iteration, no median)
• Few other outliers (spread >5%):
  • Atlas digi-reco: ~8% (1 host)
  • CMS digi: ~6% (2 hosts)
• The good news:
  • Variations in the hepscore (using Chris' hepscore.py script with the YAML input file from hep-benchmark-suite, iterations=1) are almost less than ~1%
    • 2 outliers: ~2.3%
       
• Issues with hep-benchmark-suite (run_hep-score_example.sh):
  • Tried to run on host with 2x 64core processors (256 logical processors), but it crashed because it couldn't fork enough parallel threads
    • hepscore.py works well, so I assume an issue with the hep-benchmark-suite
  • Hard to debug because of data loss - does the suite remove old folders???