Cluster Computing and JCSP Networking

摘要

Hoare's algebra of Communicating Sequential Processes (CSP) enables a view of systems as layered networks of concurrent components, generating and responding to events communicated to each other through channels, barriers and other (formally defined) synchronisation primitives. The resulting image and discipline is close to hardware design and correspondingly easy to visualise, reason about, compose and scale. JCSP is a library of Java packages providing an (occam) extended version of this model that may be used alongside, or as a replacement for, the very different threads-and-monitors concurrency mechanisms built into Java. The current release (JCSP 1.0) supports concurrency within a single Java Virtual Machine (which may be multi-processor). This paper reports early experiments with JCSP.net, an extension of JCSP for the dynamic construction of CSP networks across distributed environments. The aims of JCSP.net are to simplify the construction and programming of dynamically distributed and parallel systems. It provides high-level support for CSP architectures, unifying concurrency logic within and between processors. The experiments are on some classical HPC problems, an area of work for which JCSP.net was not primarily designed. However, low overheads in the supporting infrastructure were a primary consideration ― along with an intuitive and high-level distributed programming model (based on CSP). Results reported show JCSP holding up well against ― and often exceeding - the performance obtained from existing tools such as mpiJava and IBM's TSpaces. The experimental platform was a cluster of 16 dual-processor PIII Linux machines. It is expected that future optimisations in the pipeline for the JCSP.net infrastructure will improve the results presented here. JCSP and JCSP.net were developed at the University of Kent.