Distributed Algorithm Engineering

Abstract

When one engineers distributed algorithms, some special characteristicsarise that are different from conventional (sequential or parallel)computing paradigms. These characteristics include: the need for eithera scalable real network environment or a platform supporting asimulated distributed environment; the need to incorporate asynchrony,where arbitrary asynchrony is hard, if not impossible, to implement;and the generation of ``difficult{''} input instances which is aparticular challenge. In this work, we identify some of themethodological issues required to address the above characteristics indistributed algorithm engineering and illustrate certain approaches totackle them via case studies. Our discussion begins by addressing theneed of a simulation environment and how asynchrony is incorporatedwhen experimenting with distributed algorithms. We then proceed bysuggesting two methods for generating difficult input instances fordistributed experiments, namely a game-theoretic one and another basedon simulations of adversarial arguments or lower bound proofs. We giveexamples of the experimental analysis of a pursuit-evasion protocol andof a shared memory problem in order to demonstrate these ideas. We thenaddress a particularly interesting case of conducting experiments withalgorithms for mobile computing and tackle the important issue ofmotion of, processes in this context. We discuss the two-tier principleas well as a concurrent random walks approach on an explicitrepresentation of motions in ad-hoc mobile networks, which allow atleast for average-case, analysis and measurements and may giveworst-case inputs in some cases. Finally, we discuss a useful interplaybetween theory and practice that arise in modeling attack propagationin networks.