Achieving adaptive parallelism over TCP/IP

Abstract

Many workstations forming part of local area networks around the world are idle most of the time. In an interactive environment the processor's power is mainly used to update the display and handle messages, which is very wasteful considering that this takes a very small percentage of the total processor time. Researchers have realised that this wasted computational power could in some way be used to solve problems which require huge amounts of CPU cycles to perform. The communication structure, although not necessarily efficient, is already available. Most of these workstations are networked, and nowadays most organisations have an Internet connection to the world outside. Various systems have been developed whereby a pool of jobs is distributed to idle workstations, depending on their availability. Processors may join or leave the pool according to the amount of local work being done. The term adaptive parallelism is being used to distinguish from other forms of parallelism which have fixed resources, and which have an easily measurable speedup. Most organisations requiring years of computational power are turning their attention to adaptive parallelism, since it provides a feasible solution to investing in a single large supercomputer. In this work we will analyse the currently available systems implementing adaptive parallelism. We will also be describing a new implementation of a system of this kind, one which uses TCP/IP as a communication protocol, and can thus be extended over a whole WAN such as the Internet. The system makes use of a worker and farmer suite, and specific applications will need to implement separate 'plugins'. These modules will then attach to the worker and the farmer and use them as a framework for their communication. In this sense the system is unique. However the system is very restrictive in its domain due to efficiency considerations. It has been tested on a LAN consisting of Windows NT workstations at the University of Malta and we will be presenting an empirical and analytical study of the results.