Re: second squid box for redundancy and load balancing

On Tue, Dec 21, 1999 at 09:31:12PM +0100, Henrik Nordstrom wrote:
[load-sharing]
> The more advanced way is to use a TCP loadbalancer, for example Cache
> Server Director from RadWare or a Alteon switch or similar equipment.

This is the solution we've adopted to handle the load for enforced caching
of over 25000 machines, with an Alteon AceDirector distributing the load
between a pool of UltraSparcs running Solaris.

However, we find that this causes us a few problems. The need for
persistency on certain websites such as web-based email services (all
requests in a session must appear to come from the same IP address to avoid
authentication problems) required us to distribute the load using a hash
based on destination IP address.

This works, save that at times the load balancing between machines becomes
very skewed. In the worst case, around 70% of the traffic on one day was
going through one machine (at the time one of five), but with no sign of
this being due to a large proportion of the total traffic being to one or
two hosts. At peak times this would lead to a severe bottleneck at one or
two cache machines while the others were relatively lightly loaded.

Is this a problem that has been seen by anyone else using a similar set-up?
We submitted this problem to Alteon's technical support some time ago, but
have not as yet received a response that has resolved the matter.

-- 
--------------- Robin Stevens  <robin.stevens@oucs.ox.ac.uk> -----------------
Oxford University Computing Services  http://www-astro.physics.ox.ac.uk/~rejs/
  (+44)(0)1865: 273212 (work) 726796 (home) 273275 (fax)  Pager: 0839 629682