Transparent parallel replication of logically partitioned databases

This paper presents a protocol for efficient transaction management in an environment of replicated autonomous databases. Each replicated copy has ownership over mutually exclusive portions of the database. The protocol improves response time and throughput by exploiting parallelism although reducing the degree of transaction isolation. Most modifications to the database are assumed to be on the locally owned portion of the database, with only occasional nonlocal writes/updates. Read operations, however can access either local or nonlocal objects equally. We are able to prove that users of our parallel replicated database system can view it equivalent to that of a single database providing "degree 2" transaction isolation, i.e. the replication and parallelism is transparent to the application programmer. The protocol communication overheads are limited allowing it to be efficiently implemented over even wide area networks. Experimental results using a prototype demonstrating the performance improvements are presented.