Tweaking gossip algorithms for computations in large-scale networks
Stefan Dulman, Eric Pauwels · 2015
Gossip algorithms have already been identified as possible solutions for information aggregation in large-scale distributed systems. For example, the “classical” algorithm allows computation of the average of the values stored in a network, in a fully distributed manner. Smart extensions use this basic algorithm as a primitive for computing complex statistics of the values in the network, perform online optimization, system identification, etc. In this paper we bring into discussion a less used variant of gossip (SFC) which employs order statistics on exponential random variables to compute the sum of the values in a network. We study the two algorithms in parallel as they are in fact interchangeable. We show that for large-scale networks, SFC proves to be a surprisingly fast and robust aggregation method, easily extendable to allow self-stabilization. As the second contribution of the paper, we introduce such a mechanism and analyze its performance.