On Sun, Jan 31, 1999 at 11:13:27PM -0800, Curt Sampson said:
> On Sun, 31 Jan 1999, Greywolf wrote:
> 
> > Is any of this useful on non-symmetric disks, i.e, I have differently sized
> > partitions on differently sized disks.
> 
> I should think so. As I understand it, if the partitions are still
> different sizes, the ccd will still be allocated as _interleave_
> blocks from the first, followed by _interleave_ blocks from the
> second, and so on. Once a partition runs out, I assume it's just
> removed from the list, so with a 1GB and 1.2GB partition, you will
> get the first 2 GB interleaved, and the last 200 MB at the end of
> the 1.2 GB disk.

That's correct. I read through the CCD code when I first started playing with
it. It does the obvious thing. (It doesn't keep a list, BTW, that would be
silly. It just does the math, which is not that complicated.)

> This would still display the same problems WRT even numbers of
> _interleave_, since nothing's changed as far as which disk it's
> going to hit for a particular block on the ccd.
> 
> I expect the asymmetry is happening because the inode blocks end
> up concentrated on one disk when a power-of-two interleave is used.
> This might affect even small interleaves, because I'd bet that ffs
> uses the inodes in a cylinder group in order from the beginning.
> 
> It'd be nice to have a detailed analysis of just what changes you
> can expect to see on a ccd given different interleaves.

Actually, if anyone is going to go to the trouble of doing this, I'd be far
more interested in RAIDframe results. I jsut spent a pile of money on
another SCSI-to-SCSI RAID controller, and I'd love to skip that the next
time around. The other thing worth measuring, which I didn't do for ccd
but which should be more interesting with RAIDframe, is latency. How much
of a penalty do the extra calculations in the driver exact?

/a