Abstract
This work concerns the design and prototype implementation of an agent-based parallel architecture for physically distributed systems. The generic goal is to combine the processing power of widely available, low-cost networks of workstations, providing parallelism inside single applications. The specific goal is to investigate ways of implementing agent-based parallel processing in distributed systems. In this context, an agent is a lightweight mobile process that can freely move in the network and execute when it reaches a processing node. The Swarm architecture addresses these points by providing an abstract environment that can span many or all machines in the network. The environment is structured as a virtual machine, whose organisation and instruction set are detailed. Swarm is based on the idea of process flow, in which mobile concurrent processes can move and execute asynchronously in a distributed space consisting of data nodes. Each node is capable of permanently storing arbitrary information and references to other nodes, permitting the creation of persistent and distributed data structures in the environment. The main advantage is a flexible programming environment, which combines characteristics of the message-passing and distributed shared-memory approaches. A subset of the Swarm architecture was implemented as a prototype, coded in C language for operation under the Unix environment, to study and evaluate the model. The prototype executed in a single workstation, simulating the Swarm abstract environment and pennitting the validation of the proposed architecture and implemented mechanisms. Both the implementation and the evaluation procedure are explained and discussed. Results suggest that agent-based processing is feasible in moderately-and tightly-coupled environments, and that the Swarm processing model can be successfully applied to local-area networks and massively parallel computing machines. In particular, applications that manipulate irregular and distributed data structures can benefit from the programming environment provided by the Swarm architecture. These comprise: symbolic processing (artificial intelligence and expert systems), distributed simulation, distributed databases, and intelligent networks.