|
We were able to fix the rebuild problem by re-init'ing the whole "0+1"
disk set (8 rz29's) and having the customer restore their software to
the new logical "0+1" disk. I think the old configuration had some data
corruption that prevented the rebuild from completing.
Also thru much testing I was able to figure out how the swxcr
configures the data in a RAID 0+1 set. It work as follows.
The primary blocks of data are in capital letters (ie A,B,C..) and the
mirrored blocks are in small letters (ie. a,b,c ...) and the number of
disks in the RAID 0+1 set does not make any difference to how the data
is written on the RAID 0+1 set.
For example:
Disk 1 2 3 4 5 6 7
AgHnO aBhIoP bCiJpQ cDjKqR dEkLrS eFlMsT fGmNtu
|
| Your example has 7 disks, an odd number. RAID 0+1 ought to always have
an even number during normal operations.
For maximum availability, mirror pairs ought not to reside in the same
shelf or at least be on different SCSI buses.
Once the proper mirror pairs are established then the striping can be
configured.
For 8 disks there will be 4 mirror pairs spread across however many
shelves there are. Suppose there are 3 shelves, the most the SWXCR can
support. As long as the disks are spread out 4, 4, 0; 4, 3, 1; 4, 2,
2; or 3, 3, 2; i.e. no shelf with 5 or more disks, then the mirror
pairs can be distributed in such a way so as to avoid having the member
disks on the same shelf/SCSI bus. After that the striping will spread
chunk-sized pieces of the LBN range across the 4 virtual devices.
|