Java server scaling solution

stackoverflow! We are developing system, which should be horizontal scalable. So, messaging system seems to be the right approach, but it very low-level. Our main requirement is persistent connections between clients and server system (clients are mobile applications communicating with server by xml-based protocol). The n开发者_Go百科ext very important task is work distribution based on node current load. Now we are using legacy application based on Apache Mina framework, but it is not scalable. So, what architecture will be sufficient and what libraries or frameworks do you know to solve our problems?

Work distribution should be based on task length, which could be variable. Every application server in cluster should be able to send message to client at any time without request from client (push).

And how about Hazelcast or GridGain?