Priority {IO} Scheduling in the Cloud

Abstract

Current state of the art runtime systems, built for managing cloud environments, almost always assume resource sharing among multiple users and applications. In large part, these runtime systems rely on functionalities of the node-local operating systems to divide the local resources among the applications that share a node. While OSes usually achieve good resource sharing by creating distinct application-level domains across CPUs and DRAM, managing the IO bandwidth is a complex task due to lack of communication between the host and IO device. In our work we focus on controlling the hard disk drive (HDD) IO bandwidth available to user-level applications in a cloud environment. We introduce prioritybased (PBS) IO scheduling, where the ordering of IO commands is decided cooperatively by the host and IO device. We implemented our scheduling policies in the Linux storage stack and Hadoop Distributed File System. Initial results show that in a cloud environment, the realtime commands managed by PBS outperform the realtime IO scheduling of the Linux kernel by up to a factor of ∼ 5 for the worst case latency, and by more than 2x for average latency.