The only bit I can contribute is making it message driven vs request driven. Push instead of poll. Poll systems have the added overheard of the balance between lag between requests and requesting too much.

In a very controlled, well administered messaging system, where jobs are "sent" out to others, you'll get uncanny, natural scaling vs a pull system where you're always tuning the pulling.

Polling is great for detecting when things happen with no sense of predictability, such as figuring out if a system was completely unplugged.

Consider that if you wish to create queue systems by hand.

- segue here -

Try putting permutations of those terms into google. You may find other systems as well. Tibco is supposed to be super awesome, but.. pricey.