Crazy ZFS ZIL options: md(4) umass(4)

[Fabian Keil]

grarpamp <grarpamp at gmail.com> wrote:

> > ZIL is there for you to recover transactions in case of a crash.
> > It is your safety net.
> 
> I always thought the ZIL was pushed out safely. So that still
> no matter what the disk would be consist [1]. Like when crash
> you just lose the ZIL's since the last ZIL push. Which odds
> are will be just work product here.

I'd expect the pool to remain consistent as long as one of
the uberblocks "works", of course you'd still lose the
transactions that haven't made it to the disk yet.

I agree that using RAM that isn't battery-backed for the ZIL
doesn't make much sense, though, and that disabling sync
is more reasonable if you can live with losing transactions.

> > Use the RAM for ARC, it will provide more performance.
> 
> But about reducing fragmentation without separate ZIL.
> I'm admittedly over full and will need to move data to
> new pool anyway. Just that with ZIL in main pool what article
> I read says problem can mostly come back without separate zil.
> I tend to run full till annoyed to redesign, bad habit.

Disabling sync potentially reduces the fragmentation and you could
additionally increase vfs.zfs.txg.synctime_ms and vfs.zfs.txg.timeout
which (again potentially) reduce fragmentation further but can
negatively impact "interactivity".

I'm not aware of a quick way to measure fragmentation on ZFS
pools, though, so I'd be interested to know how you intend
to confirm that your "fragmentation tuning" actually improves
things.
 
> > USB ... unreliable
> 
> I would have to test USB bus and devs for stablility. But
> I do have local lifetime warranty on 32GiB devices :)
> So maybe mirror 2 of them, or 2x2.
> Due to crypto I only get 7-15MiB/s on spindle anyway.

I'm sure a slow ZIL could throttle this even further.

Fabian
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 196 bytes
Desc: not available
URL: <http://lists.freebsd.org/pipermail/freebsd-fs/attachments/20130215/70c327d4/attachment.sig>