Testing of a Novel Distributed Shared Memory Framework
TCD-CS-1999-67.pdf (PDF) 275.7Kb
A novel object oriented framework for software Distributed Shared Memory (DSM) has been developed by the Distributed Systems Group in Trinity College, Dublin for programming parallel applications on a group of loosely coupled workstations. It offers the developer the ability to apply different rules for how accesses to shared data are seen by each process (consistency models) which can be implemented by different means of managing shared data (coherence protocols). Different combinations of models and protocols can be applied to individual objects within an application, in addition the protocols that the framework uses can be customized by the developer. By providing this range of facilities it is hoped that developers will be able to exploit application specific semantics to achieve improvements in performance and speed-up characteristics while maintaining relative ease of programming. The objective of this project is to test to what extent this DSM framework delivers the flexibility, customizability and programmability in design of parallel applications that it promises and whether it can offer appropriate speedup characteristics for different classes of problem. Two applications with significantly different data sharing characteristics were designed and implemented as sample cases to test the framework: the Travelling Sales Person (TSP) problem which has a low communication-to-computation ratio and for which workers can work independently of each other; and the LU Decomposition problem which has a very a high communication-to-computation ratio and an algorithm where workers are highly interdependent The thesis will show that the framework did achieve appropriate speed-up characteristics. It also allows significant flexibility and programmability: protocols can be easily set and changed for individual objects; the protocols themselves were customized to improve the performance of each of the applications.
Author: Markey, Laurence
Availability:Full text available