ZFS performance gains real or imaginary?

[James R. Van Artsdalen]

Matt Simerson wrote:
> I haven't benchmarked it with -HEAD  but with FreeBSD 7, using a ZFS
> mirror across two 12-disk hardware RAID arrays (Areca 1231ML) was
> significantly (not quite double) faster than using JBOD and raidz. I
> tested a few variations (four disk pools, six disk zpools, 8 disk
> zpools, etc).

A backup server is a *highly* specialized type of server.  It's likely
that data is only rarely updated, meaning that there are very few
partial parity-stripe writes for the Areca to deal with.  A database
server receiving many updates would have an entirely different pattern
of write I/O, possibly forcing many partial stripe updates.  Since ZFS
(almost?) never does partial stripe writes in a RAIDZ the performance
comparison between ZFS with JBOD and your hardware setup might change
considerably with a database workload.  Not to mention the dominance of
sequential I/O in a backup server, etc.

For a backup server ZFS has other advantages.  A client's backup server
recently ran low on space so I took over another 4x1GB enclosure and
added it to the pool with no downtime: there were a couple of large file
writes to that pool running when I arrived that were still going when I
left.

There's also the issue of cost: once SATA port multiplier support works
in FreeBSD it will be very practical to build cheap ~15TB servers for a
small business using ZFS.