On Sun, 21 Jan 2007 23:50:28 -0500
"Blair Sadewitz" <blair.sadewitz@gmail.com> wrote:

> BUFQ_PRIOCSCAN, IMHO, is superior to BUFQ_READPRIO.
> AFAIK, it is not fully implemented, but even so it can significantly
> improve interactive response time and even write throughput under
> load.
> 
> For performance tuning, one can change the number of queues
> (PRIOCSCAN_NQUEUE) as well as the integer constant priocscan_burst[].
> The number of elements in the array corresponds to the number of
> queues.  I use three (default) queues with priocscan_burst[512 128 32]
> (this was approximately the peak of a roughly inverted-U-shaped
> curve).  Both of these are
> found in cscan_init() in kern/bufq_priocscan.c.

Some benchmarks I did with dbench time ago:

http://mail-index.netbsd.org/tech-kern/2005/10/11/0002.html

-- 
http://plog.xtrarom.org/
Juan RP's blog - NetBSD/pkgsrc news in Spanish