Opinions on safety of dropbox ???

Rob Seastrom rs at seastrom.com
Mon Aug 4 14:27:45 CDT 2014


Mike O'Dell <mo at ccr.org> writes:

> The best way to do local storage is mirrored pairs of drives.
> Yes - that uses half the storage for reliability
> but it resilvers the fastest and with 4 Terabyte
> drives so stupid cheap, you want ease of maintenance.
>
> You can then do striping on top of the mirrored pairs
> ("Raid-10") and run ZFS RAID-Zx

Close, but not quite.  You definitely do NOT want to be running any
kind of RAID underneath of ZFS.  It hides disk failures from the
filesystem/os, and makes it hard to recover from checksum errors.  You
want a controller that allows you to expose the disks to the OS in
JBOD mode.  In the vernacular mode, this is called "IT Mode"
(initiator/target).

ZFS 101: A vdev is made up of one or more disks, and can be single
disk (no redundancy!), mirror, raidz[23], etc.  You can not add disks
to a vdev after the fact.  A zpool is made up of one or more vdevs.
They don't have to be the same flavor of vdev.  You can keep adding
vdevs to a zpool after the fact, but you can't *remove* a vdev after
it's added without starting over from scratch.  Writes are striped
across available vdevs, so you can improve your performance with more
vdevs (the terminal velocity of which is having vdevs consisting
entirely of mirrors).  Going in the other direction, if you don't need
high write performance, the intent log (ZIL) is a write cache and is
optional.

Resilver on a single drive failure in a raidz (or zfs mirror for that
matter) is definitely faster than on a hardware RAID, assuming it's
not full.  Why?  After all in most cases you're constrained by the
sequential write speed of the drive you're resilvering, right?

Well, that's true as far as it goes...  but the hardware RAID
controller is pretty stupid and doesn't know anything about what
sectors are actually in use by the filesystem or anything like that.
So it has to copy *everything* over.  For the sake of argument,
consider a single-vdev zpool that is 1/3 full.  That means the failed
drive is 1/3 full too, regardless of whether it's a zfs-mirror,
raidz...  and resilvering will happen in 1/3 the time that it will
take for a hardware mirror.

More here:
http://forums.freenas.org/index.php?threads/slideshow-explaining-vdev-zpool-zil-and-l2arc-for-noobs.7775/

-r



More information about the Tacos mailing list