[OpenIndiana-discuss] SMR disks

[Aneurin Price]

On 1 May 2015 at 07:13, Nick Tan <nick.tan at gmail.com> wrote:
> Has anyone tried using SMR disks with ZFS?  I bought a Seagate 8TB SMR disk
> and put it in a esata enclosure for my backups.  I found that zfs send
> would cause the disk to go offline.  My guess is that zfs send is too fast
> and fills the drive write cache.

I've been using one of those disks for about a month now with ZFS on
Linux. It's part of a set of three mirror vdevs; the initial resilver
took about two or three times as long as I would have expected for a
traditional drive (it's hard to be sure exactly how long it would have
taken with a traditional drive since my pool is a lot different
compared to my last resilver: fuller and more fragmented), which if
anything is better than I was expecting, since the sustained and
fairly random write workload of a resilver is about the worst case
scenario for an SMR drive. Since then, I've not noticed any
performance issues, though the pool does not see heavy usage, and
nearly never sees significant random writes. Overall, I've actually
been surprised at how well it's been working: I was fully prepared for
the possibility that ZFS would present a pathological use case
rendering the drive almost useless, but in practice I've had no
trouble at all.

It is possible that Linux handles drive delays in such a way as to
lower the chance of the disk going offline, although I've actually
read about *more* problems with drive timeouts with ZFS on Linux than
on Illumos (because there are multiple points in the stack that will
wait for a timeout before reporting an error up), so I'm not sure.

Currently my drive is paired with another 4TB drive, so it's
effectively 100% over-provisioned. I don't think this should make a
difference, but it's worth mentioning just in case.

Note that the random access on-disk cache (ie, non-shingled section
that acts as a persistent cache for random writes until they are
written into place) is 25GB on this drive, so if you find that the
first 25GB or so goes at great speed and then performance drops off a
cliff and the drive starts hanging for several seconds at a time, that
will be why (though I've not actually experienced any of that
worst-case behaviour myself). This paper has some experimental
analysis of SMR performance, specifically including measurements on
the Seagate Archive drives:
https://www.usenix.org/conference/fast15/technical-sessions/presentation/aghayev