Although we cannot present performance result, we are confident that the architecture will be as efficient as the underlying software allows for. Preliminary evaluations of the non-linear solver execution time indicate that it it will largely overlap the remesh time. Optimization are possible for best exploitation of the RMI mechanism; we checked that on all the machines which run the Solver (IBM SP2 and Linux PCs), and the Domain manager (Sun workstations), the Java GC is actually able to pin down arrays.
Besides completing this experiment, future work will try to generalize this experiment to a more general HPF JNI. We plan to use an open compiler infrastructure such as ADAPTOR [1] to get rid of the technical problem of compiler-dependent initialization problems. Such an interface will probably have to use some of the Java extensions designed for integrating data and task parallelism [16,15,11] in order to be able to express the semantics of asynchronous RPC as described in this paper.