William J. Knottenbelt, Soraya Zertal, Peter G. Harrison
A Distributed Lock Manager (DLM) provides distributed applications with a convenient means of synchronising their accesses to shared resources. This paper presents a performance study of three different implementation strategies for a DLM considering both the layout of the lock database (centralised or distributed) and the strategy used to assign lock masters (static or dynamic). For each implementation strategy, we develop accurate analytical models of communication cost, resource utilisation and lock request response time. The models highlight bottlenecks in the system and show clearly for what mixes ofincoming lock request types it is best to use static or dynamic lock master positioning. The analytical formulae are validated against a detailed event-driven simulation which uses realistic hardware parameters. This validation reveals a good agreement between analytical and simulation results, particularly with respect to communication cost, node and CPU utilisation, system capacity and the response time trend.
PDF of full publication (434.3 kilobytes)
GZipped Postscript of full publication (144.8 kilobytes)Information from pubs.doc.ic.ac.uk/performance-lock-management-iee.