16-bit memory, caches, and UNIX V/386

karl lehenbauer karl at ficc.uu.net
Mon Aug 21 22:53:44 AEST 1989


In article <8746 at saturn.ucsc.edu>, laurie at ucscb.UCSC.EDU (60648000) writes:
> In article <9639 at b-tech.ann-arbor.mi.us> zeeff at b-tech.ann-arbor.mi.us (Jon Zeeff) writes:
> >Does anyone have a benchmark as to what 16 bit memory does to performance
> >if you have a caching motherboard?  I'm quite interested since it's the
> >only way I can add memory.
 
> I was running a AST premium 386/20 with 2.6mb of 32 bit memory. I then added 
> a 2mb 16 bit 100 ns memory card and when running graphics programs that 
> malloc'ed HUGE data structures (1-2 mb), my box ran about 40-60 % slower. 
> I was really suprised at the diff.

But does it have a cache?  On a Mylex 386/16, 32-bit reads that are cache 
misses incur two wait states, 16-bit reads that are cache misses incur five 
wait states, cache hits are zero wait states.  They claim the cache hit rate
(cache only helps on reads; writes must still be written, although I think
writes only incur one wait state) is typically 80%.  This should mean
(in the "typical" sceniaro) that 20% of the reads take 2.5 times longer
with 16-bit memory rather than 32.  I realize benchmarks would be a lot
more useful.

I'd like a way to differentiate between the 32-bit and 16-bit memory and use
them for different purposes.  You could isolate your 16-bit memory from
Unix simply by not making its address start contiguously with the end of
your 32-bit memory.  

The only way I have figured out to make use of that without the Unix source
(or a lot of pain) is to use it as a ramdisk.  It would be really nice if you 
could get Unix to use it for disk buffers, tho' again a pain without the source.

-- 
-- uunet!ficc!karl	"Have you debugged your wolf today?"



More information about the Comp.unix.i386 mailing list