zfs, nfs and zil

[Freddie Cash]

On Mon, Mar 28, 2011 at 4:22 AM, Claus Guttesen <kometen at gmail.com> wrote:
> I've setup a server with FreeBSD 8.2 (prerelase) and patched zfs to
> ver. 28. The server has 11 disks each 2 TB in raidz2. The performance
> is very good and I've got approx. 117 MB/s on plain GB nics using
> iscsi.
>
> I'm mounting the FreeBSD-server from a couple of vmware esxi 4.1
> servers using nfs, but when there is alot of i/o the server becomes
> unresponsive, easily triggered by installing ie. ms-sql. The server
> itself is up but is not reachable from the network. When I take the
> nic down and up again connection to the network is reestablished
> (ip-wise).
>
> A friend of mine has suggested that I disable the zil. The page
> http://wiki.freebsd.org/ZFSTuningGuide says 'Disabling ZIL is not
> recommended where data consistency is required (such as database
> servers) but will not result in file system corruption.'
>
> Has anyone tried to disable zil and achieved better performance and
> still maintain a consistent filesystem?

If your disk controller has a lot of cache on it, and a battery
backup, then enabling the write cache and disabling the ZIL can be
faster, without sacrifising consistency (the write cache on the
controller acts like a ZIL).  There's several threads on the
zfs-discuss mailing list where this is discussed.

However, the better solution, and the one most recommended for those
using NFS with ZFS, is to install a small, write-optimised, SLC-based
SSD to the system as a separate log (SLOG/ZIL) device.

NFS is a very sync-heavy protocol, and having a super-fast ZIL sitting
on a separate SSD will greatly improve things.

-- 
Freddie Cash
fjwcash at gmail.com