Re: Parent and Squid...

On Tue, 5 May 1998, Henrik Nordstrom wrote:

> Alejandro Navarro wrote:
> > Is there any way to make SERVER A the tcpconnection and when he
> > is overload send some request to SERVER B ?? I mean,...like having
> > to squid servers working in parallel mode?
>
> If the two servers should be seen as one server (located at the same
> place), then the easies way is to use DNS round-robin to evenly
> distribute the clients on both servers.

Just remain aware that the DNS round-robin method is not actually load
balancing (per se.). I think it's worth quoting Albitz & Liu (from the
O'Reily DNS bible):

 "Note that this is really load sharing, not load balancing, since the
name server gives out addresses in a completely deterministic way, without
regard to the actual load or the capacity of the servers servicing the
requests."

So if you have a common base of machinery with similar performance
characteristics and similar nominal load stats, then this would probably
be an adequate solution to distributing the hit load amongst them. Note
that we are approaching an array-based/clustering solution in this
discussion, hence the mention of CARP below.

What are the alternatives? Here's a thin slice:

A traffic redirection system ala Cisco Local Director (or Cisco WCCP on
their cache-engines)?

 http://www.cisco.com/warp/public/751/lodir/index.shtml

Alteon Lvl 4 switches? (dig around in:)

 http://www.alteon.com/wpapers/sswp.html

Software solutions running on dedicated hardware

  http://www.nsrc.nus.sg/STAFF/edward/pweb.html

NetApp (http://www.netapp.com) have incorporated a (apparently
proprietary) clustering solution. From the NetApp online blurb:

 "You can build caches that are robust, shared, and distributed by
defining a cache hierarchy or cluster. You could, for example, configure a
hierarchy of parent and neighbor caches to share objects among themselves
or to route requests through firewalls. Or, you could define a group of
caches--a cluster--that balances the load of requests among the members,
thereby increasing performance."

How does a cluster differ from a hierarchy?

 "A cluster is comprised of multiple caches (cluster mates) that are
functionally treated as one cache. The primary reason to define a cluster
is for load balancing. In most cases, configuring clusters is preferable
to configuring neighbors because fewer transactions between caches are
required for request handling. However, in certain situations, defining
neighbors might be useful for low bandwidth Internet Service Providers."

And then there's Inktomi ... and so on.

For the little guy, rolling out software on cheap Linux/BSD-on-Intel boxes
will remain the mainstay of their survival, right?

The next generation of CARP implemented on some small/large cache arrays
could be interesting. Is anyone aware of any developments on this front?

J.

----------------------------------------------------
John V Cougar | Voice: 1800 065 744
Cache Manager |-----------------------------
Telstra Internet | E-Mail: cougar@telstra.net
----------------------------------------------------
Received on Tue May 05 1998 - 18:56:15 MDT