[ Wednesday, March 1, 2000 ] Christian Robottom Reis wrote:
> I've seen a lot of variation on various runs of tiotest using the same
> setup - even in single-user mode. Is this expected, and do you know why it
> happens? Is it just the effect of the buffer cache, or do we avoid using
> it?
we don't avoid using it currently. Since I can find neither a 2.2
or 2.3 that has working i386 madvise(), it could be awhile :)
> What's a decent --numruns to use, taking into evidence such
> variation? I've noticed if I use more than one I get worse numbers in
> general - this is ok?
if you don't trust numruns > 1, don't use it :)
It may be worth watching "vmstat 1" output during a run just so you
can get an idea of the memory/caching interaction that's going on.
I'd like to believe that higher numruns further reduces the effect
of memory,,, for numruns=1,2,4 my numbers come out pretty close
Dir Size BlkSz Thr# Read (CPU%) Write (CPU%) Seeks (CPU%)
----- ------ ------- ---- ------------- -------------- --------------
. 512 4096 4 6.81976 9.47% 6.95052 10.3% 164.596 1.81%
. 512 4096 4 6.72370 8.35% 6.88223 10.3% 165.602 1.84%
. 512 4096 4 6.69172 7.37% 6.83409 10.5% 169.500 1.89%
James