Subject: Varying disk geometry
To: None <tech-kern@NetBSD.ORG>
From: der Mouse <mouse@Collatz.McRCIM.McGill.EDU>
Date: 11/22/1995 07:23:23
>> The problem, as I understand it, is that the sectors-per-track
>> figure varies depending on the cylinder number.
> I've wondered a bit if the FFS architecture might not be obsolete in
> the light of the way disks now are built. Not being a file system
> performance guru, I wouldn't know...
I'm hardly a filesystem performance guru myself...but I do know that
FFS's idea of optimal block placement and cylinder groups and suchlike
is rather silly on modern disks with varying sectors per track.
Whether it's actually a performance win in spite of that, I'd have to
run tests to find out. I would guess that it is worthwhile only to the
extent that it avoids seeking; however, FFS's idea of when the disk
seeks doesn't agree with the disk's, so it will merely avoid _long_
seeks. Again, I don't know enough about disks to be sure, but the time
taken to seek can probably be modeled as a constant overhead plus some
other constant times the number of cylinders seeked past; how
worthwhile trying to avoid long seeks is then depends on the relative
sizes of these two constants. Then there's head switch delay, which
FFS takes no account of as far as I can recall. And if bits are coming
off the platter faster than they can be dumped into RAM (which I
understand is depressingly common), you might as well forget almost all
attempts at good block placement, because the bottleneck is elsewhere.
Anyone here work for a disk manufacturer?