hi,

were these iozone runs done with a filesystem or on a raw device?
if it was with a filesystem, you could also try a run with the performence
enhancements I've been working on:

ftp://ftp.netbsd.org/pub/NetBSD/misc/chs/ubc-perf/

grab the latest diff, put the new genfs_node.h file in sys/miscfs/genfs,
rebuild.  this improves filesystem performance dramatically on old, slow
machines.

-Chuck


On Wed, Aug 22, 2001 at 04:43:04PM +0900, Toru Nishimura wrote:
> Ok, folks, here is the table of figures to show how badly DECstation
> NCR MI SCSI performance was suffered.
> 
> K48     1.5K compiled at 2000-11-21
> K67     1.5K                 -12-12  
> S0      1.5S             2001-03-15
> V2      1.5V                 -06-01
> X1      1.5X                 -08-22
> 
> There were major technological reaps done between 1.5K and 1.5S.
> I point here that NetBSD/pmax has not been useful until UBC buffer
> throttling code was added by Jason T.  Figures were collected from the
> three consecutive iterations of 40MB iozone with 25MHz MAXINE running
> on an old slow Quntaum LPS 540.  The best values are choosen of each three. 
> 
>           K48           K67             S0              V2              X1
> W sec.  36.38           35.79           72.13           55.60           72.98   
> R sec.  29.79           29.83           34.96           28.61           38.27
> 
> W B/s   1,152,914       1,171,920         581,492         754,371      553,411
> R B/s   1,407,957       1,406,069       1,199,743       1,426,027    1,083,799
> 
> With the introduction of UBC write/read performance is greatly
> fluctuated by virtual memory activity.  Sometimes write performance
> is worse than the best case figure with 5% or so.  Apparenly MIPS pmap
> code must be examined and tuned for UBC to retrieve the lost performance.
> 
> Tohru Nishimura