[net.periphs] caching disk controllers for Unix

fred@mot.UUCP (Fred Christiansen) (01/04/85)

[]
anyone have experience with these?  do they really help?
   the sector cachers say you gotta do it this way because of Unix' habit of
fragmenting a file all over the place.  but, since Unix caches blocks anyway,
does sector caching help?  my guess is it would if the controllers memory
exceeded kernel's buffer cache.
   the track cachers say their scheme helps for now and will be best once
scatter/gather HW memory management is available.  the latter seems contrary
to the free list, which is fairly integral to the current Unix kernel's view
of the file system (no reason why there couldn't be another, I suppose).

guy@rlgvax.UUCP (Guy Harris) (01/06/85)

>    the sector cachers say you gotta do it this way because of Unix' habit of
> fragmenting a file all over the place....
> the latter seems contrary to the free list, which is fairly integral to
> the current Unix kernel's view of the file system (no reason why there
> couldn't be another, I suppose).

Can you say "4.2BSD"?  I thought you could!  I suspect several other variants of
the UNIX file system have also been done that use a bit map rather than a free
list; I suspect most other operating systems use bit maps as well.  The only
think I can see that a free list buys you is that if you have limited
main memory and disks big enough that significant portions of the bit map
couldn't be kept in the buffer cache, grabbing the first blcok off the free
list rather than making an effort to pick an optimally positioned block
would require fewer disk accesses (it would also require less CPU time
under any circumstances).  Since the days of 256K PDP-11s are behind us,
I don't see the free list being worth much anymore...

	Guy Harris
	{seismo,ihnp4,allegra}!rlgvax!guy

dan@ciprico.UUCP (01/10/85)

Here at ciprico we make a caching disk controller for the multibus.  SMD
interface.  I don't have any real numbers handy at the moment but caching
helps a great deal...approx. +30% performance.   And, a lot more depends on
how the caching is performed.  We did a study here to see how unix really
accessed the disk (this means requests handed to the driver).  With some
intelligent caching I'd think we could increase the performance by at least 50%.
By all means look into caching disk controllers....if you find anything
interesting I'd be interested in hearing about it.
Oh yes...the study was done on System III...512 byte file system blocks.
-------
-Dan A. Dickey		ihnp4!mmm!ciprico!dan