Subject: RE: Port benchmarks
To: Curt Sampson" , "John A. Maier <johnam@kemper.org>
From: Bill Dorsey <dorsey@lila.lila.com>
List: netbsd-ports
Date: 03/25/2000 03:29:55
Curt Sampson wrote:
>It's a 21066, not a 21064. The 21066 is very, very slow when
>accessing the memory and second level cache, and this cripples its
>performance. (A 486 accesses memory/cache close to twice as fast,
>and a Pentium about three times as fast.)
>
>If you're doing something that remains entirely in the first level
>cache, you'll notice that a 166 MHz 21066 is about as fast as a
>P133. For anything that doesn't fit, you'll find it performs on
>par with a high-end 486.
>
>You can run lmbench to get exact numbers for memory and cache access
>speed.

I ran hbench on my Multia, Sparc 5, and PWS 433A and came up with
some interesting numbers for each machine on memory bandwidth:

Multia:	 25 MB/S (166MHZ 21066)
Sparc 5:	 50 MB/S (160MHZ Fujitsu)
PWS 433A:	200 MB/S (433MHZ 21164)

These are for very large reads, where the cache sizes have no
significant effect.  It would be interesting to see how various
i386 systems fit in here.

--
Bill Dorsey