An Interposed I/O Scheduling Framework for Latency and Throughput Guarantees

Cloud storage system is becoming a trend in production environments for its economic benefits. With such architecture, storage resource is consolidated to provide multiplexing service for concurrent applications. Therefore, storage centers must be able to guarantee multi-dimensional Quality of Service for various applications. However, satisfying performance targets for each workload is challenging, because they compete for storage resource and have various performance targets in terms of throughput or latency. In this paper, we design and implement a novel scheduler, called Arbitrator, to maintain per-application performance no matter in terms of throughput or latency. In our scheduling framework, we introduce a factor to reflect how applications are sensitive to deadline missing. The scheduler employs a feedback mechanism to monitor latency guarantees and throughput allocation for each application, and compute how much applications deviate from their performance targets. Based on the estimation, Arbitrator makes the scheduling decision to achieve latency guarantee and proportional sharing of bandwidth. We implement Arbitrator in Linux kernel and evaluate its effectiveness, and the results show the scheduler has good ability to maintain satisfactory performance for applications.