[plug] OpenSolaris, ZFS, RAID-Z Comments

Daniel Pittman daniel at rimspace.net
Thu Sep 2 11:21:45 WST 2010


Tim <weirdit at gmail.com> writes:

> Shortly after my first posted, I discovered that a) 'open' solaris is now a
> little murky (and has no 2010 release), and b) BTRFS exists!

Have (c): http://en.wikipedia.org/wiki/ZFS#Linux — native Linux ZFS.
Oh, and (d), native ZFS on *BSD.

> Looking interesting. So lets expand the discussion to include BTRFS.  How's
> it perform?

Wildly mixed, and the tuning required for various workloads is very
different.  Recent LKML reports put the default "two copy metadata" policy at
slightly slower than ext4, which is slower than ext3, in turn ext2.

None of that is a real shock, of course.  Single copy metadata apparently
improves performance for that poster to a little above ext4 but below ext3.

> For RAID10 do you still do software raid then BTRFS over the top, or does it
> do it itself similar to RAIDZ?

Internally, generally speaking, although many reports suggest that this is
slower than having BTRFS on RAID.  Less flexible, however.


Anyway, in addition to that the RHEL senior engineer assigned to review BTRFS
for their systems[1] had serious concerns about the design leading to
unbounded fragmentation (which Chris Mason and others dispute), and unarguably
found a nasty bug leading to a pathological fragmentation case (now fixed).

Plus, the thing is new and barely has a stable on-disk format.


So, um, to my mind?  BTRFS is really, really interesting (and so is native
ZFS, since that has ... design risks too).  In about five years when other
people finish losing data to edge cases, when performance stabilizes, and when
it is clear that the design isn't broken at heart I look forward to deploying
it.

Until then XFS will do me: a very long and stable history, high performance,
and a record of both scaling up to multi-core and multi-node NUMA systems, and
of eliminating performance bottlenecks in the file system as time goes on.

        Daniel

Footnotes: 
[1]  ...at least in Europe; I don't know how Edward Shishkin fits in globally
     to their organization.

-- 
✣ Daniel Pittman            ✉ daniel at rimspace.net            ☎ +61 401 155 707
               ♽ made with 100 percent post-consumer electrons



More information about the plug mailing list