[lustre-discuss] Question about PFL

[Andreas Dilger]

In Lustre 2.16.0 there is a configurable parameter to allow a "soft" disabling of setstripe behavior, see LU-17628 "lfs: add lfs_setstripe admin restrict".

There are also OST Pool Quotas that can be used to limit usage of OSTs on a per-user, per-group, or per-project basis.  So if you have e.g. some research groups that can have a directory with the flash OST pool, but others do not, then even if they bypass the "lfs setstripe" restriction, they would not be able to use capacity in flash OST pool.  OST Pools and Pool Quotas have some limitations on how they can be used, but are generally sufficient for most sites.

Cheers, Andreas

On May 12, 2025, at 23:55, Noskov, Dr. Sergey <noskov at uni-mainz.de> wrote:

Hello Andreas,

Hello Andreas,
Thank you for your detailed answer. I think this really brings clarity to our problem.
Because most of the files are output logs from SLURM jobs etc. that could actually be generated as O_APPEND.
A suggestion for further development would be that in Lustre such a mechanism as PFL for administrators, but not changeable in the user space. This would be very handy for us if we have one metadata structure but different OSTs for different departments.

Best regards

Sergey Noskov



Von: Andreas Dilger <adilger at ddn.com<mailto:adilger at ddn.com>>
Datum: Dienstag, 13. Mai 2025 um 05:34
An: Colin Faber <cfaber at gmail.com<mailto:cfaber at gmail.com>>
Cc: Sergey Noskov <noskov at uni-mainz.de<mailto:noskov at uni-mainz.de>>, "lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>" <lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>>
Betreff: Re: [lustre-discuss] Question about PFL

Having the actual file layouts would definitely help. There is one "normal" scenario where the PFL layout will be ignored, which is for files created with "open(O_APPEND)" (typically log files, or redirected output like "command >> file" ).

For performance and scalability reasons, the append-only files are by default 1-stripe files and do not specify a pool, though this is configurable via "mdd.*.append_stripe_count" and "mdd.*.append_pool".

Recently this has come up as a point of confusion with observant users, especially with the increasing usage of PFL and multiple pools in the same filesystem.

Two things we are looking to improve in this area (LU-17159):
- mark files that use the "append" layout with a flag so that they can be identified more easily (this would appear in "lfs getstripe output" and the man page)
- use only the first component of the PFL layout for the append files if no explicit append layout is configured.  I think this would give more control to users, and be more "intuitive".

Feedback is welcome.

Cheers, Andreas


On May 12, 2025, at 06:51, Colin Faber via lustre-discuss <lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>> wrote:
You don't mention the version. Also lfs getstripe of the parent directory and one of the files would be helpful as well

On Mon, May 12, 2025, 4:18 AM Noskov, Dr. Sergey via lustre-discuss <lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>> wrote:
Hello everyone,
we have two PFL pools in the lustre.
We notice that there are files that are in the directory of pool1, but for some reason they belong to pool2.
Would that be possible that PFL does not always work?

With best regards

Sergey Noskov


_______________________________________________
lustre-discuss mailing list
lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>
http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org
_______________________________________________
lustre-discuss mailing list
lustre-discuss at lists.lustre.org<mailto:lustre-discuss at lists.lustre.org>
http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org

Cheers, Andreas
—
Andreas Dilger
Lustre Principal Architect
Whamcloud/DDN




-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20250514/2d31ed38/attachment-0001.htm>