RE: [squid-users] ICP link load-balancing... without ICP?

G'day squid-users,

In case anyone is interested, we are now running ICP <-> ICMP ping mappers
on our upstream proxies that don't support ICP. We are using this to load
balance four links to different parent proxies, two that do support ICP and
two that don't. The links are all of different bandwidths and have different
additional traffic on them. The load balancing is working nicely, with the
ICP ping times of all parents sitting below 200ms most of the time.

> -----Original Message-----
> From: BAARDA, Don [SMTP:don.baarda@baesystems.com]
> Sent: Thursday, July 05, 2001 2:00 PM
> To: 'squid-users@squid-cache.org'
> Subject: [squid-users] ICP link load-balancing... without ICP?
>
> G'day squid-users,
>
> We have several upstream parent proxies on different links. Our links are
> low-bandwidth, and the upstream proxies have heaps of bandwidth. Hence,
> hit
> or miss, it doesn't really matter which upstream proxy we use, what
> matters
> most is we use the one on the least loaded link.
>
> We have been using ICP to nicely load-balance the links, combined with a
> low-ish ICP timeout and round-robin on all the parents. This meant that we
> used ICP to load balance until all parents exceeded the ICP timeout, then
> we
> fall back to round-robin. The beauty of this is we load-balance until all
> links are saturated, then use round-robin to minimise the latency.
>
> The problem is, some of the ba@#$#ard upstream proxies have started to
> deny
> or block ICP. One is using an icp_deny acl because we are getting
> ICP_DENIED
> responses, and the other seems to have outright firewalled it. They don't
> seem to be prepared to be cooperative about it. I've tried using the UDP
> echo port trick, but it's turned off too. Interestingly, our Squid doesn't
> fetch for an ICP_DENIED response, but it still marks that parent as up and
> uses it when our Squid falls back to round-robin.
>
> I'm at the point where I'm about to write an ICP to ICMP converter so that
> our Squid's ICP requests go out to the peers as ICMP and get converted on
> the way back into ICP_MISS. Before I go through the effort of doing this,
> has any one else got any thoughts, ideas whatever on a solution?
>
> ABO
Received on Sun Jul 08 2001 - 19:20:05 MDT