AWS meeting 2023 06 08

EESSI/AWS sync meeting (2023-06-08)

• link to AWS project doc: https://docs.google.com/document/d/1CHG9fCh2LkfJ-EI8J-_Wr5NpHL5iwm8Wu6syfK9h7-c

Next meeting

• Thu 13 July 2023, 12:00 UTC

Notes previous meetings

• 11 May 2023: https://github.com/EESSI/meetings/wiki/AWS-meeting-2023-05-11
• 13 April 2023: https://github.com/EESSI/meetings/wiki/AWS-meeting-2023-04-13
• 9 Mar 2023: https://github.com/EESSI/meetings/wiki/AWS-meeting-2023-03-09
• 11 Jan 2023: https://github.com/EESSI/meetings/wiki/AWS-meeting-2023-01-11

Notes 8 June 2023 (12:00 UTC)

• status update on sponsored credits
  • Costs are about $3k/month for March-May'23 (up from ~$1.5k/month)
    • EFS costs cut in half after cleanup
    • $15k left under end of Nov'23
  • Is there a tutorial about how to explore and manage costs?
    • Should we have alarms for everything? Is this possible?
    • Need to get familiar with Cost Explorer
      • Allows cost by tags - allow us to classify things
        • tags already there for CitC Slurm cluster (fair-mastodon)
      • EBS is the thing that may cost us in the end
      • Current EFS cost is about 800USD a month (used to be double that - before cleanup)
        • ~1.8TB in size (only for Slurm cluster, CVMFS Stratum-1 is using 1 TB dedicated disk)
        • should look into tiering in our EFS filesystem, to offload stuff not being access to cold storage - see https://aws.amazon.com/efs/features/infrequent-access/ + https://aws.amazon.com/blogs/aws/new-amazon-efs-intelligent-tiering-optimizes-costs-for-workloads-with-changing-access-patterns/
      • New S3 POSIX plugin that may be interesting
• Looking into using a CDN
  • partially motivated by really low bandwidth when talking to CVMFS Stratum-1 mirror server
  • even different CDNs for different contents (CloudFront in AWS vs CloudFlare, Azure CDN, ...)
  • could help with costs, bandwidth, ...
  • observed ~40MB/sec when having both CVMFS mirror and ParallelCluster in eu-west
    • vs 2MB/sec when having ParallelCluster in eu-central (and CMVFS mirror in eu-west)
  • what speedup would we get when adopting a CDN?
• Tech short on EESSI
  • Edit looks good
  • Tweaked some of our links to new domain (eessi.io, eessi.io/docs)
  • Integration with ParallelCluster (separate video?)
    • Situation with libfabric is a bit tricky but solveable
      • We control the linker, so there are tricks we can use
      • However, have noticed issues that need discussion (demo prepared)
    • GPU support is in the works
      • Proof of concept working
      • Permission from NVIDIA to distribute CUDA compat libraries under CUDA EULA
      • Need to figure out how to integrate things with our linker
        • Relevant libraries are listed in https://github.com/apptainer/apptainer/blob/main/etc/nvliblist.conf
        • We want a specific order for library resolution (need to handle compat libraries, host libraries and containers)
        • We control the linker in EESSI so this is do-able