On Fri, Jun 25, 1999 at 01:12:11AM +0900, HEO SeonMeyong wrote:
> 	Hi. I'm HEO SeonMeyong.
> 
> 	I'm very sorry for late reply.
> 
> > Now that I'm thinking on it, are you *sure* it is IDE-related ?
> > I'm sitting in fron of a SCSI-only box rigth now, and it feels slow as well
> > when doing a I/O (a 'make depend' in /usr/src in my case).
> > Could someone confirm the same behavior with a scsi box ?
> 
> 	Now I think this probrem (NetBSD becomes more slowly) is not
> 	related IDE, because my other SCSI-only machine becomes more
> 	slowly.
> 
> 	This week I'm very busy, then I cannot run bonnie on
> 	1.4-Release/1.3I installed machine.
> 	I will test bonnie, and other benchmarks at next week or next next 
> 	week.

I actually think this is I/O related, maybe something to do with the
buffer cache and I/O queues. My 'make depend' doesn't eat much CPU
(apps are still responsive as before when they need CPU), but an I/O
takes a long time. So anything that needs to read or write a file on
disk gets blocked, although top says there's a lot of CPU idle.
My theory is that the requests stay in the I/O queue for too long, because
resquest from 'make depend' gets prepended and not appended. Or maybe it's
that buffers we just read are recycled immediatly, discarding effect of
read-aehad ?

Does someone remember of a change that affected the way the I/O queue is
managed ? Maybe something in disksort() ?

--
Manuel Bouyer, LIP6, Universite Paris VI.           Manuel.Bouyer@lip6.fr
--