gdt@ir.bbn.com (Greg Troxel) writes:
> I believe rotational optimization is mythical due to whole-track
> buffering (as well as not knowing the layout), but staying 'close' in
> cylinder space probably still helps, in terms of small seeks vs
> whole-disk seeks.

I think another important optimization may be bunk too -- that the
only thing wrenching the disk head back and forth is the code for that
filesystem.  We haven't had single partition spindles since before I
started admining.  I just don't see how the disk arm is going to be
where the FS code thinks it will be anyway.  Any "elevator" type
algorithm that sorts the per-partition motion for "closeness" is going
to clash with the same code running on behalf of the other fs
partitions (or even swap) on that disk.

(This assumes my understanding of how things works is correct.  I
*think* the I/O sorting is all on a per-partition basis.)

I've started making one-partition per spindle client machines, but
can't bring myself to make a 200 GByte "root" partition on my main
server.  The thought of a half-hour fsck that puts all my data at risk
(every time the kernel sneezes on a reboot) has me a bit concerned.

-wolfgang
-- 
Wolfgang S. Rupprecht 		     http://www.wsrcc.com/wolfgang/