A Stochastic Process Approach to Model Distributed Computing on Complex Networks

Abstract

In this work we present analytic expressions for the expected values of the performance metrics of parallel applications when the distributed computing infrastructure has a complex topology. Through active probing tests we analyse the structure of a real distributed computing environment. From the resulting network we both validate the analytic expressions and explore the performance metrics under different conditions through Monte Carlo simulations. In particular we gauge computing paradigms with different hierarchical structures in computing services. Fully decentralised (i.e., peer-to-peer) envi- ronments provide the best performance. Moreover, we show that it is possible to improve significantly the parallel efficiency by implementing more intelligent configurations of computing services and task allocation strategies (e.g., by using a betweenness centrality measure). We qualitatively reproduce results of previous works and provide closed-form solutions that link topology, application’s structure and allocation parameters when job dependencies and a complex network structure are considered.