IT-CIS post-AFS usage (EOS/CEPH/S3/..)

Wednesday 19 Nov 2014, 16:00 → 17:00 Europe/Zurich

CERN

present: Luca, Andreas, Jiri, Tibor, Pedro, JeanYves, Samuele, Jan

Reason for the meeting: recent instability in AFS (power cut exposed the fact that only user homedirectories are currently on critical power; AFS client crashes sees on some machines; kernel-level "hang" in AFS client for some unavailable volume might tie up threads in the middleware), observe IT-CIS interest in non-AFS storage (ongoing discussions Tim/Massimo, some SNOW tickets). IT-DSS would like to make sure that their storage services are aligned with any new requirements, and that common solutions are investigated with preference.

INDICO: (not much changed since last meeting)

• AFS dependencies: still looking at readonly volumes for 'top of the tree'
• 2016 (or end of 2015) expect to migrate to some form of object store

OpenDATA: (currently on EOS, ~30TB)

• would need backup (worried about eventually being the single remaining data repository)
  • EOSPUBLIC will get a tool to backup (readonly) parts into CASTOR, prototyped with CMS. Could give early access via EOSPPS but given the close general availabiliy this is not required
• currently throttling to 10..20 concurrent connections on their side. Should not really be an issue on EOSPUBLIC, can increase
• prefer to access data via HTTP (at least for readonly; writes/copy-in are done out-of-band via "eos" tools)

CDS:

• on AFS (3M docs, 1TB), would like something else. Can use "eos" or "xrdcp" for writing, prefer read access via HTTP.
• incoming: digitized photographies
  • video collection is still on DFS (easier to access from transcoding machines)
• current CDS software needs a (small but highly-availble) shared filesystem where session data is written to disk, then some worker node reads that based on a shared DB (i.e. no directory listing, no "state machine"). Currently on AFS. Expect to switch to REDIS in 6months, but would prefer something on critical power for the interim time.