The Multiple Pool Migrating Worker Paradigm : A Distributed System Framework and Model

Abstract

A network of workstations (NOW) can provide an inexpensive and effective distributed processing platform. The purpose of this thesis is two-fold, providing a methodology for distributed computing on a NOW first, and providing a model to predict and monitor performance second. The Multiple Pool-Migrating Worker (MPMW) paradigm uses multiple job pools to divide up tasks and migrating workers to balance the work load. The MPMW paradigm is a quick and efficient way of implementing problems using distributed processing without extensive knowledge of parallel programming. A model describing the MPMW paradigm is developed using queuing theory and Mean Value Analysis techniques. The model connects run time, granularity and scalability. It is designed to allow prediction of run time from system measurables. The experimental results show that the model describes a system implementing the MPMW paradigm and allows estimation of resources such as the run time in systems with multiple processors. The paradigm and its model are used to examine the task granularity verses scalability trade-offs in a system. The experimental application used is a serial LU decomposition. There are many overhead costs involved in using a NOW for distributed processing, the more workstations, the more overhead, the less performance. However, the results show that the granularity of the tasks in the job pools can affect system performance. It is further shown that scalability can be manipulated by changing the granularity of the job pools. Both task size and bundle size affect performance in distributed systems and the MPMW model presented in this thesis predicts this general behavior. However, the impact of non-optimal task sizes is minimized with migrating workers and job pool bundles.