[Jlab-scicomp-briefs] Nov 19: Transition of the Farm /cache filesystem to read-only from farm and ifarm

[Wesley Moore]

Transition of the Farm /cache filesystem to read-only from farm and ifarm

What is changing?
On November 19, 2024, during monthly maintenance, the /cache filesystem will be changed to read-only access from farm, ifarm, Globus, and XRootD Data Transfer Nodes. After that date, files will only be copied to /cache via in the following ways:

  *
The jcache command, which reads files from tape and writes them into /cache
  *
The jput command, which writes files to tape and can optionally place a copy immediately in /cache when the -cache flag is specified
  *
Data ingest from the experimental halls using the jmirror command with a regular expression pattern match for data retention in cache.


Why is this happening?

  *
The new system will ensure that files move to tape promptly and that /cache is an accurate subset of files stored on tape.
  *
In the current /cache filesystem, there are a commonly cases where files are in conflict with tape storage, leading to work slow downs.
  *
Small file handling has been a historic problem, and many small files in /cache were not stored on tape or backed up in any way


How will this affect farm job workflows?

  *
Jobs that are part of a SWIF workflow with an output specification to /cache will continue to work.
  *
Jobs that attempt to write directly to cache using open(), cp, mv, or other POSIX tools will fail.  Output from slurm jobs that are not part a SWIF workflow should be stored on /volatile and will need to be moved to tape manually using jput on ifarm.  Generally, slurm workflows that need to interact with tape would be better implemented using SWIF.
  *
Note that jput is not available on (non-interactive) farm nodes because it may queue, stalling the farm node and potentially timing out the job.


What is not changing?

  *
Cache deletion policy: remains unchanged.
  *
Cache file pinning: continues to work as before.
  *
jcache client: continues to work as before.
  *   SWIF outputs to /cache: continues to work as before.


References:

  *
KBA: Migration to read-only cache<https://jlab.servicenowservices.com/scicomp?id=kb_article_view&sysparm_article=KB0015468>
  *
KBA: Computing Coordinators<https://jlab.servicenowservices.com/scicomp?id=kb_article_view&sysparm_article=KB0014686>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.jlab.org/pipermail/jlab-scicomp-briefs/attachments/20241028/244469ff/attachment.html>