Details
-
Improvement
-
Resolution: Fixed
-
Minor
-
Lustre 2.16.0
-
3
-
9223372036854775807
Description
The patch https://review.whamcloud.com/35034 "LU-12273 lod: metadata overstriping" allows an MDT to hold more than one stripe of a directory, in order to raise the maximum directory size limit, as well as increase parallelism in directory locking at both the LDLM and OSD level.
However, this introduces a potential incompatibility with older clients that don't understand overstriped directories. It would be good to add a tunable parameter to limit the usage of this feature. It would be named like "lod.*.max_stripes_per_mdt" (or "mdt_max_overstripe_factor" or similar), and replace most uses of the hard-coded constant LMV_MAX_STRIPES_PER_MDT in the code. On master for 2.16, lod_max_stripes_per_mdt would be initialized to LMV_MAX_STRIPES_PER_MDT=5 (the hard upper limit due to transaction sizes), and if this is backported to older branches for interoperability it would initialized to "1".
In lod_prep_md_striped_create() if a new directory was requested with stripes_per_mdt > lod_max_stripes_per_mdt, then IMHO it would be better to silently clamp this value to lod_max_stripes_per_mdt to avoid application failure. If lod_max_stripes_per_mdt == 1 then the existing check in lod_mdt_alloc_rr() will strip the LMV_HASH_FLAG_OVERSTRIPED flag from the layout during creation, to maintain compatibility with older clients.
