Exploiting concurrency to improve latency and throughput in a hybrid storage system

Abstract

This paper considers the problem of how to improve the performance of hybrid storage system employing solid state disks and hard disk drives. We utilize both initial block allocation as well as migration to reach “Wardrop equilibrium”, in which the response times of different devices equalize. We show that such a policy allows adaptive load balancing across devices of different performance. We also show that such a policy exploits parallelism in the storage system effectively to improve throughput and latency simultaneously. We implemented a prototype in Linux and evaluated it in multiple workloads and multiple configurations. The results show that the proposed approach improved both the latency of requests and the throughput significantly, and it adapted to different configurations of the system under different workloads.