Optimistic replication of data is becoming increasingly popular in distributed systems. However, the performance characteristics of such systems are poorly understood. We have developed a generalized simulation framework to investigate the operating costs and service quality of optimistic replication for file systems. We describe the simulation and present the results of a validated, trace-based simulation study of an optimistic replicated file system.

The simulation can support a large-scale, heterogeneous, distributed environment specified by its own input language. The abstraction of library components allows the extension of our simulator to match a variety of system settings. The input traffic emulates actual trace data collected at a representative office-computing environment. The simulator is developed using Maisie, a C-based discrete-event simulation language, which enables parallel execution of the simulation.

Our simulation revealed that the widely used conflict rate metric is misleading: If you never exchange updates, you will never have a conflict. We investigated this anomaly and proposed alternative metrics to evaluate the service quality of optimistic replication. In addition, the deduced operational costs from simulation suggest that we can economically achieve a high quality of optimistic replication service. We demonstrated also that optimistic replication scales well, and the low percentage of active users in a large distributed system self-regulates service quality. Finally, we discovered the interdependence of frequency, method, and path of data propagation, on the operating costs and service quality of optimistically replicated file systems.

Back to Main Page (frame disabled)
Last modified: June 25, 1998