Research On Key Issues Of Performance Optimization Of A Cluster File System

Improving the performance of cluster file system is an ever-lasting researchtopic in I/O for Cluster architecture. This paper presents the author's research onthree key issues in performance optimization of COSMOS, a cluster file system forDawning-3000 super-server, including client-side caching, scalable storage serviceand performance evaluation.In Chapter 1, this paper introduces the basic concepts and features of clusterarchitecture and cluster file system first, then it gives an introduction of related workin this area. The methods to improve the scalability of cluster file system are alsoreviewed in this chapter. In Chapter 2, the background for this paper is discussed.Chapter 3 turns to Client-side Caching. The author presents his thoughts oncooperative cache and coherence protocol first, then gives a detailed description ofcache policy used in COSMOS with its implementation and evaluation.Chapter 4 focuses on the scheme of scalable storage service in COSMOS,including its distributed storage and control architecture and a method to enhancethe performance of storage server. Our preliminary performance testing results provethat this scheme can improve the scalability of COSMOS greatly.Finally, this paper concentrates on performance evaluation for general clusterfile systems. The author refines current evaluation methods here. Then he setups theperformance evaluation model for read/write and aggregate bandwidth testing forthose cluster file system employed with storage stripes, and illustrates the possibleways and limitations for further performance optimization. At the end of Chapter 5,a method to evaluate the throughput of small file operations is introduced with itstesting results in COSMOS.