An-I Andy Wang
Florida State University
Geoff H. Kuenning
Harvey Mudd College
UCLA, The Laboratory for Advanced Systems Research
Optimistic replication provides high data availability in the presence of network outages. Although widely deployed, this relaxed consistency model introduces concurrent updates, whose behavior is poorly understood due to the vast state space.
This paper introduces the notion of permuted states to eliminate system states that are redundant and unreachable, which can constitute the majority of states (4069 out of 4096 for four replicas). With the aid of permuted states, we are for the first time able to construct analytical models beyond the two-replica case. By examining the analysis for 2 to 4 replicas, we can demystify the process of forming identical conflicts-the most common conflict type at high replication factors. Additionally, we have automated and optimized the generation of permuted states, which allows us to explore higher replication factors (up to 10 replicas) using hybrid techniques. It also allows us to validate our results with existing simulations based on actual replication mechanisms, which previously were analytically validated with only one pair of replicas.
Finally, we have discovered that update locality and bimodal access patterns are the primary factors contributing to the formation of identical conflicts.