Improving Performance in Replicated Databases through Relaxed Coherency

Rainer Gallersdörfer, Matthias Nicola

           Applications in finance and telecommunications (intelligent
           network, network management, mobile computing) cause renewed
           interest in distributed and replicated data management. Since
           synchronous update of replicated data is experienced to
           degrade distributed systems performance substantially,
           relaxing the requirement of coherency (mutual consistency)
           has become a favorable approach to achieve high performance
           in replicated databases. In this paper we present formal
           concepts for specifying relaxed coherency which allows
           to calculate an independent measure of relaxation, called
           coherency index. We incorporate this metric into a detailed
           analytical queueing model which emphasizes on the quality of
           replication to evaluate the impact of relaxed coherency on
           the performance of replicated databases. The model considers
           response time, throughput, scalability and network traffic as
           performance criteria. As it turns out, performance improvements
           through relaxed coherency depend significantly on various
           system parameters. We closely examine the trade off between
           consistency and performance, and show that in many situations
           a slight relaxation of coherency can increase performance
           remarkably. Finally, we verify these results by benchmarking
           an implementation of relaxed coherency.