Load balancing for minimizing the average response time of get operations in distributed key-value stores

Abstract

We investigate the impact of an unevenly distributed load among nodes of a distributed key-value store on response times. We find that response times of “get” operations quickly degrade in the presence of power law distributions of load and identify the point, at which the system needs to apply a mitigation approach. The migration technique, which we propose, overcomes the long response times of consistent hashing placement techniques. Our technique is a hybrid approach that combines consistent hashing and a directory for exceptions. Our experimental results show an improvement in the average response times and an equal load among the nodes.