Local high performance networks availability already makes workstations clusters a serious alternative for parallel computing. However a high level and effective programming language for such architecture is still missing. Recent works show the interest in Java for cluster programming. One of the main issues is to handle efficiently the communication of objects to really take advantage of the network speed. The paper presents an alternative to the standard serialization process through the proposal of a Java object communication library. Object allocation is controlled in such a way that the transfer of objects between two nodes comes to a direct memory to memory dump. We show how specific allocation mechanisms can cooperate with a Java Virtual Machine so that fast transfers of graphs of objects can be achieved. Experimental results are given for basic operations and for a genetic programming application; they demonstrate a dramatic change in the transfer speed