Broadcast-Based Parallel Programming

The properties of a parallel programming model are largely determined by the model of communication/interaction upon which it is based. Parallel programming models, languages and implementations have historically been formulated in terms of shared memory and/or message passing communication models. This paper motivates parallel programming models based on broadcast communication and defines and illustrates a parallel programming system based on associative broadcast. Associative broadcast is a dynamic multicast where the set of recipients for a message is dependent on the state of the processes in the broadcast network and is determined at runtime. Associative broadcast supports direct implementation of fully distributed (peer to peer) control. It enables dynamic composition of programs from components by integrating discovery, linking and invocation. A programming language and runtime system implementing the associative broadcast programming model are discussed including example programs and results. Parallel programs formulated in this programming model are intrinsically dynamically structured and faulttolerance via replication is readily obtained.