Topic
2 replies Latest Post - ‏2012-09-19T12:52:40Z by SystemAdmin
SystemAdmin
SystemAdmin
2092 Posts
ACCEPTED ANSWER

Pinned topic GPFS performance with full or nearly full filesystenm

‏2012-09-18T23:49:56Z |
When we have a filesystem on GPFS become full or almost full, all of GPFS - including other filesystems - slows to a crawl. Apparently this is a well-known phenomenon, and the only way to avoid it is to keep the filesystems from getting full. General folklore seems to say that filesystems should not get under 90 percent. This slowdown seems to happen quite abruptly - it doesn't come on gradually as the filesystem fills up. Is there some way to know where this threshold is? If it can't be calculated beforehand, is there some way to check to see whether it is in "filesystem-full slowdown mode." Or even better, is there a way to keep this from happening in the first place?
Updated on 2012-09-19T12:52:40Z at 2012-09-19T12:52:40Z by SystemAdmin
  • Tucks
    Tucks
    78 Posts
    ACCEPTED ANSWER

    Re: GPFS performance with full or nearly full filesystenm

    ‏2012-09-19T09:42:16Z  in response to SystemAdmin
    We don't see this at all. We've filled our file systems up to 99.9999% many times (Linux NSDs with DS3512 storage).
    Are you running scattered or clustered?

    To keep it from happening, use HSM, delete data or use a big stick on your users.

    For performance, possible mmpmon I suspect (others chime in here).
    • SystemAdmin
      SystemAdmin
      2092 Posts
      ACCEPTED ANSWER

      Re: GPFS performance with full or nearly full filesystenm

      ‏2012-09-19T12:52:40Z  in response to Tucks
      I have two clusters - one with three filesystems of 100 TB each, 200 clients, and scatter allocation; and another with 30 filesystems varying from 4 TB to 10 TB, 400 clients, and cluster allocation. The effect is more pronounced on the first one.