[squid-users] Starvation by diskd

I'm working with a FreeBSD-based gateway machine which is serving
as a NAT router, recursive DNS resolver, and VPN server. It's also
running Squid as a transparent proxy.

The system was seeing frequent dropped VPN connections, especially
under heavy network loads.

After looking the system over and trying a few experiments, it
looks to me as if VPN tasks (pptpd and ppp) and other networking
tasks are being starved for CPU cycles by diskd. It may also be
affecting natd (FreeBSD's NAT daemon, which runs in userland at
standard priority) and named.

I've niced Squid down by 1 (to 1 below the standard priority, so
that it has lower priority than natd or named) and have niced diskd
down to standard priority, and the disconnections have stopped. I
haven't changed the priority of unlinkd yet, but am thinking that
it might be best to change it to one level above standard priority.

Are there "knobs" in squid.conf to control the priorities at which
Squid, diskd, and unlinkd run? Shouldn't there be? And given that
diskd apparently polls for I/O completion on many systems,
shouldn't it have a lower default priority?

--Brett Glass
Received on Tue Sep 26 2006 - 10:47:42 MDT