Modeling and control of distributed asynchronous computations

Abstract

A stochastic model for a class of distributed asynchronous fixed point algorithms is presented and a methodology for optimizing the rate of convergence is introduced. An important parameter in the authors model, called the degree of synchronization, quantifies the average amount of time each processor is willing to wait for information from other processors (before beginning computation of its update variable based on the available estimates of variables from other processors). The authors analyze the relationship between the convergence rate and the degree of synchronization for a class of iterative fixed point algorithms. Preliminary analysis indicates that significant improvements in convergence rates can be achieved by proper control of the parameters in the authors model.<>