RE: [squid-users] squid loadbalancing cluster -- best practice

> > Hi Everyone,
> >
> > Nowadays I am planning building a load-balanced squid + web nodes
cluster.
> > The detailed information:
> > Operating System: Linux 2.6.9
> > Squid-2.6-Stable-15
>
> I'd highly recommend upgrading your requirements a little:
>
> Linux 2.6.22+ (Ideally 2.6.24, but that may still be too cutting-edge for
> high-performance use, and 2.6.20 is the minimum for a number of apps)
>
> Squid 2.6.stable19
>

On this point, I will follow your advice this time.

> >
> > Here we have two options:
> >
> > squidA + Apache1
> > Lvs +
> >
> > squidB + Apache2
> >
> > (Lvs = linux virtual server)
> > SquidA forms sibling relationship with SquidB using ICP.
> > Apache1 is parent (original server) to SquidA.
> > Apache2 is parent (original server) to SquidB.
> >
> >
> > The other is
> >
> >
> > Squid-parentA + apahce1
> > Squid-child +
> > Squid-parentB + apahce2
> >
> >
> > Squild-child round robins Squid-parentA and Squid-parentB.
> >
> > Anybody has any suggestions on the two options?
> > Which is better of the two implementations?
>
> That depends on your expected traffic volumes and speed requirements. I
> have a vague idea what you mean by LVS. You mean using DNS to round-robin
> the A-record IPs and just point at a CNAME right?

LVS is a very good Open Source Linux Load-balancing Software by Wensong
Zhang, and it is proven to be very efficient and stable in large websites.
For your reference, http://www.linuxvirtualserver.org/
Using this LVS and High-Arability Suite, such as heartbeat we can avoid the
single point of failure problem.

> For highest performance and reliability I'd pick the first option unless
> there was some other topology feature that obsoletes the reliability edge.
>
> Performance-wise;
> current squid production releases are proven only to around 800req/sec.
> So below that either setup is equally effective. But above that the
> first proposal will be better for you.
>
> Reliability-wise;
> The second config leaves you with a single-point bottleneck which may at
> some point fail. If it does it takes out access to the child proxies.

>

In the end, I appreciate your great suggestions and will take on some
upgrades and experiments.

Xu Feng
Shanghai,China
Received on Mon Apr 07 2008 - 23:30:11 MDT