On May 15, 10:28 pm, cjt <
chelj...@prodigy.net> wrote:
> Application: primarily a High-Def video server, but also some
> miscellaneous related and unrelated storage. The typical video file is
> 2-4 GB, but some are as large as 40 GB.
>
> About 75% of the video is quasi-archival (seldom watched); the rest can
> be pretty active (multiple simultaneous streams). The whole thing is
> read-mostly.
>
> Some possibilities I've considered:
>
> Three concatenated banks of eight drives, with each bank either raidz or
> raids
>
> Six banks of four, each bank raidz
>
> Four banks of six, raidz or raidz2
>
> I like the idea of using raid rather than mirroring (despite the
> decreased performance) because of the economy. I do want to have
> something that I can later expand in reasonable increments while
> preserving some semblance of order. I haven't quantified the needed
> performance beyond a gut feel that I need decent, but not exceptional,
> performance. The current server is nearly out of capacity and can
> sometimes barely keep up -- it's on considerably slower hardware, with
> 12 drives in a single raidz2 (I've since read that having that many
> drives in a single raidz2 isn't a good idea, so I'm hoping to do better).
>
> All insights are appreciated.
This workload is a good candidate for a RAIDZ configuration, but
another really important consideration is your ability to backup
your data. If you can't back it up, then don't build it.
You could go with a smaller RAIDZ2 config and then add another
RAIDZ2 vdev. Keep in mind that you can't attach disks to your
existing RAIDZ vdevs, you can only add another vdev.
I would not go lower than RAIDZ2 with a large pool like this one.
You could start with 4 vdevs of 4 disks, then add another 4 disk
vdev with 4 spares left. Or, consider a smaller mirrored pool for
your non-video data.
Mirrored pools provide good performance and are more flexible
than RAIDZ configs.
You might check out the ZFS best practices, here:
http://docs.oracle.com/cd/E23823_01/html/819-5461/practice-1.html
Thanks,
Cindy
\