Re: [squid-users] multiple squid 3.1.10 instances - independent vs frontend/backend

From: Amos Jeffries <squid3_at_treenet.co.nz>
Date: Sat, 08 Jan 2011 13:22:36 +1300

On 08/01/11 07:56, guest01 wrote:
> Hi guys,
>
> I am using a couple of squid instances per server (Squid 3.1.0, RHEL
> 5.5, lots of RAM) and was wondering which would be the better
> configuration? Better in that case means more performance.
>
> Either
> 1) - 4 separate squid instances, each with its own completely
> independent cache? (no cache hierarchy or something like that)
> or
> 2) - a configuration similar to [1] with frontend and backend instances?
>
> Does anybody have experiences with either of these two configurations?
> We are currently using configuration #1 with 20 RPS/per instance and
> load values of up to 8. It is not much, but we are authenticating with

That load value is high, Squid-3.1 *should* be handling a several
hundred RPS easily on modern CPUs.

> Kerberos/NTLM/LDAP and even content filtering by ICAP. Load is pretty
> high, but since we are using 16 CPUs per server, it is ok until up to
> 16. CPU usage stays pretty low (<20%), and disk IO is not an issue
> (high IO idle values, very low IO wait values) too, I am wondering why
> the load is that high.

ICAP by its nature adds a lot of overhead. As does NTLM. Avoiding either
is a big speed boost. 3.1 is a little CPU hungry, we are working on that
and some small memory issues which remain.

NP: ICAP has some cutting-edge features not yet in 3.1 which allow
bypassing ICAP and trickle-feeding for services which support it. This
reduces the ICAP overheads and will be out with 3.2 some time next year.

> What do you think, could configuration #2 improve performance (testing
> it would be best, but it is not that easy with a live system)?

It is easy with a live system. You pick which type of instance (front or
back) you need to add. Implement it. Then push some small portion of the
traffic through it and increase until no more gain is seen (usually
around 50Mbps and some lucky people have recently reported 200Mbps).

The backends of a (2) system are individual instances of (1), so you can
implement (1) and then add instances for (2) to see if the gain is worth
it. ExtremeCarpFrontened wiki page covers how to integrate those.

Amos

-- 
Please be using
   Current Stable Squid 2.7.STABLE9 or 3.1.10
   Beta testers wanted for 3.2.0.4
Received on Sat Jan 08 2011 - 00:22:41 MST

This archive was generated by hypermail 2.2.0 : Sat Jan 08 2011 - 12:00:02 MST