[zfs-discuss] ZFSoL Performance

Gregor Kopka zfs-discuss at kopka.net
Thu Feb 4 01:23:34 EST 2016


Yes, dedup can kill your performance, especially on deletes and rewrites.

> tank  dedupratio                  1.62x                       -
>
> tank  free                        30.3T                       -
>
> tank  allocated                   13.2T         

In case you can live with short downtimes you still have a way out:

First add some SSD as L2ARC, disable dedup (to stop it from getting
worse) and enable secondarycache=metadata (to have it cached in L2ARC)
on all datasets, then iteratively for all datasets:
zfs snapshot, send/recv to a new dataset, repeat send with -i till you
are current, initiate downtime, snapshot and send -i again to get the
last changes, move the new dataset into place of the old (by
manipulating mountpoint, rename, whatever works), end downtime, then
drop the old dataset (which will give you the dedup performance penalty
because it has to rewrite all DDT related to that dataset, but hopefully
the DDT will mostly be in ARC and L2ARC at that time - do not reboot to
not lose L2ARC contents).
This will effectively remove dedup from the pool.

As a remark I think that raidz1 is, while giving a good cost/space
ratio, to low in redundancy with 4TB drives because of very long
resilver times.

Gregor

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.zfsonlinux.org/pipermail/zfs-discuss/attachments/20160204/e24d1a42/attachment.html>


More information about the zfs-discuss mailing list