On Wed, Jun 19, 2002 at 05:00:56PM -0400, Perry E. Metzger wrote:
> 
> It certainly isn't a reliable way to find out these stats. vmstat -s
> will do better I believe.
> 
> >From what you've said, though, you have execmin set to 5, which means
> on a 128 meg machine like yours you'll be miserable. Set it up to 20
> and let us know how things feel.

It's not a tuning problem. I looked at this as well, and couldn't find a
tuning that makes it behaves properly. Running X, while doing a cp
(a dd if=/dev/zero of=file bs=64k is enouth, in fact), the X server is
stuck. A top running remotely (text mode is fine, the problem is really with
the X server) shows that X is blocked on biowait. No need for memory-ungry
apps, just X with a xterm is enouth.

While copying to another disk than the system disk, the system behaves
properly. However a systat shows activity on the system disk (small but
constant) that should not be there (the machine is doing nothing but
running top in a xterm). It starts as soon as the write start, and stops
as soon as the write stops. I don't know yet what cause this (if it's
activity to swap partition, or if it's some file that the X system has to read
for each event, or such).

Anyway, I really think the rigth way to fix this is not to bend the VM
system to handle X especially, but to different I/O priorities, based
on process I/O behavior. This would solve a whole class of problems.

-- 
Manuel Bouyer <bouyer@antioche.eu.org>
--