A Stochastic Analysis of the Performance of Distributed Databases With Site and Link Failures

Abstract

A stochastic model for analyzing the performance of a distributed database is proposed. The database is prone to site and link failures, possible leading to a partition of the underlying communication network. The system model is parametrized to support very general assumptions about data replication, transaction access patterns and network connectivity. For concreteness of analysis, a concurrency control protocol based on Thomas' Majority Consensus protocol and the Two-Phase Commit Protocol is used. A new performance measure called expected system degradation is proposed; this measure is a combination of availability of data and the transaction response time; this is the first step towards the ultimate goal of defining the notion of availability for real-time transaction systems. The model allows a database designer to analyse the expected system performance and choose the right input parameters that emphasize the relative importance of availability and response times.