On 05/02/2010 04:11, Edward Ned Harvey wrote:
Data in raidz2 is striped so that it is split across multiple disks.
Partial truth.
Yes, the data is on more than one disk, but it's a parity hash, requiring
computation overhead and a write operation on each and every disk. It's not
simply striped. Whenever you read or write, you need to access all the
disks (or a bunch of 'em) and use compute cycles to generate the actual data
stream. I don't know enough about the underlying methods of calculating and
distributing everything to say intelligently *why*, but I know this:
Well, that's not entirely true. When reading from raidz2 (non-degraded)
you don't need to re-compute any hashes except for a standard fs block
checksum which zfs checks regardless of underlying redundancy.
In this (sequential) sense it is faster than a single disk.
Whenever I benchmark raid5 versus a mirror, the mirror is always faster.
Noticeably and measurably faster, as in 50% to 4x faster. (50% for a single
disk mirror versus a 6-disk raid5, and 4x faster for a stripe of mirrors, 6
disks with the capacity of 3, versus a 6-disk raid5.) Granted, I'm talking
about raid5 and not raidz. There is possibly a difference there, but I
don't think so.
Actually, there is.
One difference is that when writing to a raid-z{1|2} pool compared to
raid-10 pool you should get better throughput if at least 4 drives are
used. Basically it is due to the fact that in RAID-10 the maximum you
can get in terms of write throughput is a total aggregated throughput of
half the number of used disks and only assuming there are no other
bottlenecks between the OS and disks especially as you need to take into
account that you are double the bandwidth requirements due to mirroring.
In case of RAID-Zn you have some extra overhead for writing additional
checksum but other than that you should get a write throughput closer to
of T-N (where N is a RAID-Z level) instead of T/2 in RAID-10.
See
http://milek.blogspot.com/2006/04/software-raid-5-faster-tha_114588672235104990.html
--
Robert Milkowski
http://milek.blogspot.com
_______________________________________________
zfs-discuss mailing list
zfs-discuss@opensolaris.org
http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
