John,
I agree that if your problem space consists of routable
proxies with strong client affinity, then that is what you should
address first (while keeping more general load balancing demands in
mind, iff possible).
My understanding is that we will start discussing OPES
rechartering very soon. Adding "OPES metadata exchange" protocol into
the discussion mix would be nice. Ideally, for the discussion to be
effective, there has to be a draft with a good outline of what the
problem space and possible solution directions are. Active
participation of interested parties in the WG would be key as well.
Thank you,
Alex.
On Mon, 12 Apr 2004, John G. Waclawsky wrote:
To get back to the discussion. It seems that many of the assumptions
in the background of our discussions about load balancers are based
on web load balancers that provide virtual addresses and hide the
server cluster from the user. But, I have been thinking a little
differently in considering usage of an opes framework in the mobile
wireless market segment. This segment may be unique in that load
balancers do not rebalance for every client connection. Instead they
typically rebalance on a "per client" basis. This is done because
the devices that are being balanced have client information
associated with their traffic such as $, access rights, browser form
factors, XML style sheets...etc. Affinity to a particular server is
strongly maintained for each client. Once affinity is in place the
load balancer could operate, for example, by just replacing the
destination MAC address as the traffic arrives based on the source
IP address (you could view the load balancer as just a "bump on the
wire"). I believe this additional information should help clarify
what I am thinking, a bit. I guess I am looking for the simplest
solution for this environment and I keep thinking the most elegant
solution would be lower layer one.... just some of my thoughts...
It seems that of the two limitations you suggest, the best way to
proceed (IMO) would be to assume proxies with exposed/routable
adresses. I am thinking this assumption is the best one because it
probably provides a faster solution, would be more flexible, and
satisfies small deployment needs that do not have load balancers but
still need the metadata returned to a specific server. I think we
are also in agreement that with either limitation we still need
something to support the metadata information exchange. I was also
wondering if anyone else has any suggestions or additional
information to consider?
Regards John