My guess is you would only need a bunch of sockets, not threads or program instances. Keep track of the load on each socket and distribute it accordingly.
For my current single outgoing socket implementation I keep a list of the timestamp on each sent message. When a timestamp is older than 30 seconds I can discard it. If my list is less than 100 (or whatever limit) timestamps long I know I can send another message. I intend to scale this by adding sockets/connections to spare that will be used when the primary is above limit.