Iterative algorithms in a data-driven environment

Abstract

Data-flow principles are applied to a numerically intensive application: the Jacobi method for solving linear systems. A modification to the algorithm that enables full use of the parallelism inherent in the method by 'vectorizing' a portion of the calculation and allowing some amount of 'look-ahead' in the termination criterion is introduced. Resource allocation issues are considered, and by a combination of analytical and simulation methods, a priority mechanism which provides both an increase in performance and better resource utilization is demonstrated.