An-I Andy Wang and
UCLA, The Laboratory for Advanced Systems Research
Rajive L. Bagrodia
UCLA, Parallel Computing Laboratory
Optimistic replication of data is becoming increasingly popular in distributed systems. However, the performance characteristics of such systems are not well understood. We have developed a simulation to evaluate optimistically replicated file systems in terms of their operating costs (e.g., computing resources) and service quality (e.g., consistency of data perceived by users).
The simulation built-in language and library components ease the specification of various system settings. The traffic generator captures the inverse relationship between sharing and write accesses, and the traffic is based on a three-month long trace collected at Locus Computing Corporation.
Our result shows that the widely used conflict rate metric is misleading: Both extremes of data propagation frequencies result in near zero conflict rates. Also, conflict resolutions can generate many intermediate versions of data that are in conflicts, or meta- conflicts. We have proposed alternative service quality metrics and ways to reduce meta-conflicts.