Object-oriented implementation of data-parallelism on global networks

An alternative approach to the SPMD implementation of data parallelism is described. In this approach, the data-parallel program is executed on a collection of processors that can be located anywhere on a global network. On each of these processors a server is installed that hosts one or more servants - objects intended for performing parallel operations. One of the servers, the conductor, plays a special role: it executes the whole program except for operations on mapped arrays; to execute such operations, the conductor employs the servants mentioned above. The conductor executes on a chosen processor, the host. In addition to the parallel execution model, a framework is described that makes possible experiments with remotely executed data-parallel programs. This framework combines a High Performance Fortran (HPF) subset with the infrastructure of CORBA. It is implemented in Java. Occasionally, we quote the kernel of the algorithm generating communication sets, which was designed for the framework's implementation.