Communication Patterns for Randomized Algorithms

Abstract

Examples of large scale networks include the Internet, peer-to-peer networks, parallel computing systems, cloud computing systems, sensor networks, and social networks. Efficient dissemination of information in large networks such as these is a funda- mental problem. In many scenarios the gathering of information by a centralised controller can be impractical. When designing and analysing distributed algorithms we must consider the limitations imposed by the heterogeneity of devices in the networks. Devices may have limited computational ability or space. This makes randomised algorithms attractive solutions. Randomised algorithms can often be simpler and easier to implement than their deterministic counterparts. This thesis analyses the effect of communication patterns on the performance of distributed randomised algorithms. We study randomized algorithms with application to three different areas.

Firstly, we study a generalization of the balls-into-bins game. Balls into bins games have been used to analyse randomised load balancing. Under the Greedy[d] allocation scheme each ball queries the load of d random bins and is then allocated to the least loaded of them. We consider an infinite, parallel setting where expectedly λn balls are allocated in parallel according to the Greedy[d] allocation scheme in to n bins and subsequently each non-empty bin removes a ball. Our results show that for d = 1,2, the Greedy[d] allocation scheme is self-stabilizing and that in any round the maximum system load for high arrival rates is exponentially smaller for d = 2 compared to d = 1 (w.h.p).

Secondly, we introduce protocols that solve the plurality consensus problem on arbitrary graphs for arbitrarily small bias. Typically, protocols depend heavily on the employed communication mechanism. Our protocols are based on an interest- ing relationship between plurality consensus and distributed load balancing. This relationship allows us to design protocols that are both time and space efficient and generalize the state of the art for a large range of problem parameters.

Finally, we investigate the effect of restricting the communication of the classical PULL algorithm for randomised rumour spreading. Rumour spreading (broadcast) is a fundamental task in distributed computing. Under the classical PULL algo- rithm, a node with the rumour that receives multiple requests is able to respond to all of them in a given round. Our model restricts nodes such that they can re- spond to at most one request per round. Our results show that the restricted PULL algorithm is optimal for several graph classes such as complete graphs, expanders, random graphs and several Cayley graphs.