Actor-Oriented Programming for Resource Constrained Multiprocessor Networks on Chip

Multiprocessor Networks on Chip (MPNoCs) are an attractive architecture for integrated circuits as they can benefit from the improved performance of ever smaller transistors but are not severely constrained by the poor performance of global on-chip wires. As the number of processors increases it becomes ever more expensive to provide coherent shared memory but this is a foundational assumption of thread-level parallelism. Threaded models of concurrency cannot efficiently address architectures where shared memory is not coherent or does not exist.

In this thesis an extended actor oriented programming model is proposed to enable the design of complex and general purpose software for highly parallel and decentralised multiprocessor architectures. This model requires the encapsulation of an execution context and state into isolated Machines which may only initiate communication with one another via explicitly named channels. An emphasis on message passing and strong isolation of computation encourages application structures that are congruent with the nature of non-shared memory multiprocessors, and the model also avoids creating dependences on specific hardware topologies.

A realisation of the model called Machine Java is presented to demonstrate the applicability of the model to a general purpose programming language. Applications designed with this framework are shown to be capable of scaling to large numbers of processors and remain independent of the hardware targets. Through the use of an efficient compilation technique, Machine Java is demonstrated to be portable across several architectures and viable even in the highly constrained context of an FPGA hosted MPNoC.