Decentralized distributed computing system for privacy-preserving combined classifiers - Modeling and optimization

Abstract

The growing amount of various kinds of information triggers the need to develop efficient network computing systems, as single machines in many cases are not able to provide effective processing and analysis. One of the very promising approaches of distributed data analysis is combined classification, which could be relatively easily implemented in distributed computing systems. In this paper we address problem of decentralized distributed computing system for mentioned above classification method. We focus on the system fairness. The performance metric is defined as a maximum response time, i.e., the computing system should be designed to minimize the response time of each client using the system. We assume that the system is decentralized and each request is sent by the client directly to computing nodes without assistance of a central service. An ILP (Integer Linear Programming) model is formulated and applied to obtain optimal results provided by branch-and-cut algorithm included in the CPLEX solver. Widespread simulations are performed to evaluate properties of the computing system in terms of several parameters describing the system. © 2011 Springer-Verlag.