meeting Oct 6 2022

Notes for 2022-10-06 meeting

• date & time: Thu 6 Oct 2022 - 14:00 CEST (12:00 UTC)
  • (every first Thursday of the month)
• venue: (online, see mail for meeting link, or ask in Slack)
• agenda:
  • Quick introduction by new people
  • EESSI-related meetings in last month (incl. EESSI Community Meeting)
  • Progress update per EESSI layer (incl. bot for software layer)
  • archdetect alternative to archspec
  • 2021.12 version of pilot repository + outlook to next pilot version
  • AWS/Azure sponsorship update + OCRE funding opportunity
  • Update on MultiXscale EU project
  • Past & upcoming events
  • Q&A

Slides

• https://github.com/EESSI/meetings/blob/main/meetings/EESSI_meeting_20221006.pdf

Meeting notes

(by Kenneth)

Quick introduction by new people

• Danilo González (HPCNow!)
• Massimiliano Culpo - archspec + Spack developer/maintainer

EESSI-related meetings in last month

(see slides)

Progress update per EESSI layer

Filesystem layer

(see slides)

Compatibility layer

(see slides)

Software layer

(see slides)

• slight change of tactics on GPU support
  • script is available in PR#172 to strip down a CUDA installation to only the runtime libraries

Build-and-deploy bot

(see slides)

archdetect alternative to archspec

• archdetect pure bash is being implemented as an alternative to archspec
• different reasons to consider an alternative to archspec:
  • archspec does way more than what we need, like specifying compiler flags to use for a specific CPU microarchitecture
  • this aspect also slows down PRs that make archspec aware of new CPUs, which is unfortunate from our perspective
  • archspec releases are relatively slow, which is also annoying from the EESSI point of view
  • some CPUs are not "correctly" detect, like Ampere Altra being detected as "graviton2", due to use of product names vs actual CPU microarchitecture (neoverse-v1)
    • see https://github.com/archspec/archspec-json/issues/41, but also https://github.com/archspec/archspec-json/issues/38 (Milan recognized as zen2 because some CPU features are not exposed)
    • we may want to discriminate between AMD Milan and Milan-X, based on different cache sizes, because compiling and using different binaries for CPUs with a larger cache may help
• a pure bash implementation helps to avoid Python overhead
• archdetect would be easier to maintain (from EESSI perspective), and we would be in control over when changes are made

2021.12 version of pilot repository + outlook to next pilot repository

(see slides)

AWS/Azure sponsored credits

(see slides)

MultiXscale EU project

(see slides)

• expected start date: 1 Jan 2023

EESSI Community Meeting

(see slides)

Q&A

(see slides)