Brian Densmore wrote:

>Rather than using behemoth hardware, what about a "load-balancing"
>server situation. Something sort of like network load-balancing. You
>would need some kind of monitoring system, so that no one server would
>get more connections than it could handle comfortably.
>
>Something like this (pardon my poor graphics):
>____________    ____________       ____________
>| client 1 |    | client 2 |  ...  | client n |
>|__________|    |__________|       |__________|
>     |               |                  |       _____
>----------------------------- ... --------------|lan|
>           ___|______                           |___|
>           | Load    |
>           |balancer |
>           |_________|
>                |
>   ----------------------------- ... ----------------
>       |                |                  |
>   ____|_______    _____|______       _____|______
>   | server 1 |    | server 2 |  ...  | server m |
>   |__________|    |__________|       |__________|
>
>

Thus far, when our customers have requested >50 users, we have suggested
either more than one server and a segmented network OR a behemoth w/
gigabit backbone. That will always be prefered over something like the
above diagram because it decreases the number of points of failure. The
above suggested diagram is not far from what we're planning for other
cases but the trouble has been that you don't just "have a table that
lists all the servers with a number of max clients, keep a running tally
of current clients, and direct traffic from the client to the proper
server". For one thing, all X11 communications are UDP which, as far as
I know, is impossible to automatically get around an IP NAT. At the
kernel level, Linux has something thats relatively new in the IPv4 stuff
that allows for transparent balancing as you suggested but we haven't
tried it yet as there has been no demand.

Also, on clusters, there is a pretty high level of hardware requirements
you have to meet before doing a cluster becomes cost effective. Some of
the hardware thats available now-a-days in rack configurations is really
impressive; the sky is literally the limit on some of that stuff. We've
tersly examined the possiblity of doing a cluster and believe it would
be possible but haven't explored it.