Early work by the current authors on Java MPI bindings is reported in [2]. A comparable approach to creating full Java MPI interfaces has been taken by Getov and Mintchev [16, 10]. A subset of MPI is implemented in the DOGMA system for Java-based parallel programming [12, 13]. A pure Java implementation of MPI built on top of JPVM has been described in [6]. So far these systems have not attempted to use object serialization for data marshalling.
For an extensive discussion of performance issues surrounding object serialization see section 3 of [11] and references therein. Work of the Karlsruhe group is also reported in [17]. The discussion there mainly relates to serialization in the context of fast RMI implementations. As we may anticipate, the cost of serialization is an even more critical issue in MPI, because the message-passing paradigm usually has lower overheads.