Research consistency and perfomance of nosql replicated databases

Abstract

This paper evaluates performance of distributed fault-tolerant computer systems and replicated NoSQL databases and studies the impact of data consistency on performance and throughput on the example of a three-replicated Cassandra cluster. The paper presents results of heavy-load testing (benchmarking) of Cassandra cluster’s read and write performance which replicas were deployed on Amazon EC2 cloud. The presented quantitative results show how different consistency settings affect the performance of a Cassandra cluster under different workloads considering two deployment scenarios: when all cluster replicas are located in the sane data center, and when they are geographically distributed across different data centers (i.e. Amazon availability zones). We propose a new method of minimizing Cassandra response time while ensuring strong data consistency which is based on optimization of consistency settings depending on the current workload and the proportion between read and write operations.