An Analysis of the Impact of Network Configuration on the Reliability of Distributed Systems

One of the advantages of distributed systems is their capability to improve the accessibility of data. Fault tolerance of data is improved by replication, where data is stored redundantly at multiple sites. This increases the likelihood that the data remains accessible in spite of site and communication failures. As networks grow and evolve, parts of the network can become separated form each other into sub-networks by link communication failure for example. This creates a situation where machines in one such sub-network are able to communicate with each other, but not with machines in other parts of the network. This peculiarity called partitioning may decrease the accessibility of data. There are two techniques for measuring the accessibility of data in a distributed system: availability and reliability. Availability has received much more attention the than reliability because reliability is much more difficult and impractical to analyse. Availability of a system is important, however, there are applications where reliability is a more important measure of a system's behaviour. Since reliability is a measure of continuous availability there may be situations where a system has high availability but be quite unreliable. This dissertation describes the development of a comparative reliability measure that provides a satisfactory and practical technique for the comparison of replication control algorithms. This measure is used to investigate the effect of the network topology, consistency control techniques and copy placements on the reliability. A simple notation is developed and is used to illustrate the presentation of large amounts of comparative reliability data in an easily assimilated form which can be used to conduct performance investigation to aid analysis of how different replication techniques are affected by network configuration and copy placement