AI

Maestro: Quality-of-Service in Large Disk Arrays

Abstract

Provisioning storage in disk arrays is a difficult problem because many applications with different workload characteristics and priorities share resources provided by the array. Currently, storage in disk arrays is statically partitioned, leading to difficult choices between over-provisioning to meet peak demands and resource sharing to meet efficiency targets. In this paper, we present Maestro, a feedback controller that can manage resources on large disk arrays to provide performance differentiation among multiple applications. Maestro monitors the performance of each application and dynamically allocates the array resources so that diverse performance requirements can be met without static partitioning. It supports multiple performance metrics (e.g., latency and throughput) and application priorities so that important applications receive better performance in case of resource contention. By ensuring that high-priority applications sharing storage with other applications obtain the performance levels they require, Maestro makes it possible to use storage resources efficiently. We evaluate Maestro using both synthetic and real-world workloads on a large, commercial disk array. Our experiments indicate that Maestro can reliably adjust the allocation of disk array resources to achieve application performance targets.