Disk Tree: A Fat-Tree Based Heterogeneous Multi-tier Storage Architecture

Abstract

Traditional disk arrays have the internal bus bottleneck and can not scale well. This paper proposes Disk Tree, a novel Fat-tree based heterogeneous multi-tier storage architecture. Disk Tree is built from a collection of storage nodes, small storage appliances containing commodity disks, CPU, RAM and interconnection interfaces. Each node provides moderate levels of reliability, availability, and performance. We examine four possible configurations: striping, mirroring, logging and caching Disk Tree. We have implemented a three-tier and seven-node Disk Tree prototype and carried out extensive performance evaluations using popular benchmark for these four configurations. Depending on different I/O workloads, these four configurations perform quite differently. In general, mirroring and striping disk tree performs well on read-intensive applications while logging and caching disk tree performs well on write-intensive applications.