Does ZFS currently attempt to group metadata in large sector-ranges
on the disk? Can this be expected to happen "automagically" - i.e.
during each TXG close we have to COW-update whole branches of the
blockpointer tree, so these new blocks might "just happen" to always
coalesce into larger sector groups?
Rationale: larger regions dedicated to smallish block pointers might
be better cached (i.e. with VDEV prefetch) and speed up housekeeping
including scrubs and zfs sends. They could also reduce fragmentation
of "bulk" userdata stored in larger blocks.
These could also prove useful to implement my "Feature #1044: ZFS:
Allow specifying minimum dataset/metaslab block size AND alignment"
posted in https://www.illumos.org/issues/1044
Industry examples of grouped metadata could include copies of the
FAT table in FAT/pcfs, and the $MFT pseudofile on NTFS. Difference
is that ZFS metadata size is dynamic, so metadata regions should
not rely on any predefined sizes.
If this feature is not yet present, is it difficult or unreasonable
for some reason to add one?
zfs-discuss mailing list