Azure meeting Mar 18 2022

EESSI/Azure/SURF sync meeting 2022-03-18

Agenda

• Overview of spent credits
• EESSI hackathon January 2022
• EUM'22
• Monitoring
• Repository for hosting datasets (e.g. for WRF)
• EESSI paper published!
• Azure support for CitC
• European project proposal (funding for EESSI)

Attendees

• Bob Dröge (RUG)
• Kenneth Hoste (HPC-UGent)
• Laura Redfern (MS Azure)
• Alan O'Cais (CECAM)
• Hugo Meiland (MS Azure)
• Julian Kreuk (MS Azure)
• Ahmad Hesam (SURF)

Notes

Use of sponsored credits

• January 2022: ~€1195
  • temporary Magic Castle cluster EESSI hackathon (Jan'22): ~€610
  • AMD Milan build node: ~€350
  • Stratum-1 in us-east: ~€200
• February 2022: ~€185
• March 2022 (partial): ~€103
• in last 1.5 months: basically only Stratum-1 in us-east

EESSI hackathon

• https://github.com/EESSI/meetings/wiki/EESSI-hackathon-Jan'22
• show & tell meeting on last day to wrap up the hackathon
  • slides: https://github.com/EESSI/meetings/blob/main/meetings/EESSI_hackathon_2022-01_show_and_tell.pdf
  • recording: https://www.youtube.com/watch?v=pRjm8cayxi0
• good progress on:
  • installing software on top of EESSI (experiments with LAMMPS)
    • see https://github.com/EESSI/hackathons/tree/02_software_on_top/2022-01/02_software_on_top
  • good progress on support for running software on NVIDIA GPUs
    • see https://github.com/EESSI/hackathons/tree/05_gpu/2022-01/05_gpu
    • working solution (proof-of-concept) for Rocky Linux (should work on RHEL8-based Linux distros)
    • some form of approval from NVIDIA to ship CUDA runtime in EESSI (but nothing more than that)
    • still leaves the CUDA compilers
    • even if we include CUDA runtime, we'll still need a script to ensure things work properly
      • can also include sanity checks
    • Hugo: what would be needed in Azure VM images for this?
      • Alan: basically only reasonbly recent GPU drivers
      • script can install compat libs in userspace to fill the gaps where needed
    • scripts we have should get integrated in EESSI to facilitate letting play with it
    • Hugo also has some contacts in NVIDIA that could be helpful
    • Hugo is interested on playing with proof-of-concept GPU support with Devito (see https://github.com/easybuilders/easybuild-easyconfigs/pull/14984)
  • archiving EESSI software installations into a container image
    • see https://github.com/EESSI/hackathons/tree/16_export_software_stack/2022-01/16_export_software_stack

EUM'22

• https://easybuild.io/eum22
• incl. talk by Hugo & Davide on running WRF on Azure via EESSI
  • see https://easybuild.io/eum22/#azure-eessi-wrf
  • some additional progress has been made there
  • question in this context that came up was whether Intel oneAPI compilers and tools (incl. MKL) could be included in EESSI
  • EESSI community has not contacted Intel yet at all for this
• two talks on EESSI:
  • Getting Started with EESSI
    • see https://easybuild.io/eum22/#eessi-getting-started
  • Semi-automated workflow for adding software to EESSI
    • https://easybuild.io/eum22/#eessi-workflow

Monitoring

• good progess, work done by Terje @ Univ. of Oslo
• https://monitoring.eessi-infra.org

Data repository

• see Hugo's proposal at https://github.com/EESSI/filesystem-layer/pull/112
• WRF input data is handful of 1GB files
• should be doable with a "normal" data repository?
  • Bob will look into this
• expectation is that additional data would be pulled in soon for seismic data
• some concerns there w.r.t. client CernVM-FS cache
  • larger cache than 10GB would be needed for this

EESSI paper published

• EESSI: A cross-platform ready-to-use optimised scientific software stack
• https://doi.org/10.1002/spe.3075
• Open access
• "Software: Practice and Experience" special issue on "New Trends in HPC"

Azure support for CitC

• work done by Hugo
• resulted in PRs to CitC:
  • https://github.com/clusterinthecloud/terraform/pull/68
  • https://github.com/clusterinthecloud/ansible/pull/118
• Matt Williams (main developer of CitC) has been granted access to EESSI Azure credits for testing CitC on Azure
• maybe we should try and get Matt to join next EESSI hackathon
• Magic Castle already has Azure support, but no auto-power-down of unused workernodes there
  • Hugo could help here too
  • Alan: doing things securely is a big concern here

European project proposal

• seeking letter of support from Azure
  • if possible, also some confirmation that sponsored Azure credits could be made available to project (what we basically already have)

ISC'22

• any plans?
• half-day EasyBuild tutorial where EESSI will be shortly featured
• Laura will check if there could be a session setu p

azhop

• integration is done by Hugo, using EESSI is opt-in

Internal presentation in Azure by Hugo

• was well received
• Hugo is making it very clear that EESSI is still pre-production

Q&A

• Hugo: best option to spend time on stuff that's helpful for EESSI?
• Kenneth: blockers for stable EESSI are:
  • dedicated manpower (would be resolved if European project proposal gets accepted)
  • proper support for NVIDIA GPUs
    • Alan could post instructions on how to play with proof-of-concept that was puzzled together during last hackathon
  • GitHub App to automate workflow for community contributions
    • starting point by Kenneth: https://github.com/boegel/pyghee
  • set up central Stratum-0 server securely (physical box, yubikeys)
  • any helps on these aspects is welcome :)