Hi Jeff,

Somone from the ZFS team might want to comment, but I've been down this road
a bit so I thought I'd add my two cents.

ZFS doesn't let you create a mirror of RAID-Z vdevs on the command line, but
it is possible if you're willing to construct such a configuration
programmatically (in fact, this is what's done for the ztest test program
in usr/src/cmd/ztest/ztest.c).

An important constraint on mirrors in ZFS is that all elements have to have
identical configurations. That is, you can't have one half be RAID-Z and the
other be a raw defive or file. Unfortunately, that's exactly what you'd
have if you exported a ZVOL backed by RAID-Z over iSCSI. What you'd need to
do is export the individual disks via iSCSI and then build the RAID-Z vdev
on the data manipulating machine (the "central site that would logically
host the data"). This has some obvious deficiencies -- you're shipping parity
over the wire -- but it works.

I don't know the current state of remote replication in ZFS, but that might
provide a more efficient solution at some point.

Adam

On Thu, Oct 05, 2006 at 01:50:08PM -0700, Jeff Anderson-Lee wrote:
> We're setting up a large data store and and would like to use zfs with 
> some sort of distributed mirroring for disaster recovery.  We have 3 
> "sites" in mind: two machine rooms in one building and one in another 
> across the street, all connected by GbT networking.
> 
> Our thought was to keep 3 mirrored copies of the data, each protected by 
> Raid-5 (raid-z?).  Remote sites would export raw or raid-5/z protected 
> disk space using iscsi to the central site that would logically host the 
> data.
> 
> Given the size of the data sets (~100TB), 3-way mirroring alone is (on 
> paper at least) insufficient.  Given the number of blocks involved, the 
> chances of all three copies being bad for *some* logical block are too 
> large.  On the other hand, 6-way mirroring seems excessive.  (Our 
> original concept was to use 4-of-15 erasure coding, so that any 4 of 15 
> related blocks would be enough for disaster recovery, while each site 
> would normally hold 5 blocks, allowing for Raid-5 style local recovery 
> of single failed blocks/disks.)
> 
> Questions:
> 
> 1) is it possible to layer mirroring on top of Raid-Z under ZFS?
> 
> 2) Is is possible/practical to export raw disks or raid5/raidZ disk 
> space using iscsi?
> 
> 3) Are there some other, easier alternatives we haven't thought of?
> 
> Thanks in advance.
> 
> Jeff Anderson-Lee
> Petabyte Storage Infrastructure Project
> University of California at Berkeley
> 
> 
> 
> 
> _______________________________________________
> storage-discuss mailing list
> [email protected]
> http://mail.opensolaris.org/mailman/listinfo/storage-discuss

-- 
Adam Leventhal, Solaris Kernel Development       http://blogs.sun.com/ahl
_______________________________________________
storage-discuss mailing list
[email protected]
http://mail.opensolaris.org/mailman/listinfo/storage-discuss

Reply via email to