Research On Improving Performance And Correctness Of Cluster File System

Computer storage system, especially the slowest peripheral storage system has beenthe bottleneck which heavily influenced the performance of computer system. Currently, thecluster system has become the mainstream architecture of supercomputers and superservers,but peripheral storage system still played an important role stunting the improvement ofthe performance and stability of modern computer systems. With the high speeddevelopment of network technology in recent years, the cluster filesystem based on thehigh-speed interconnect network has been one of most efficient approaches to resolve theso-called I/O bottleneck.However in cluster system, applications express different requirements in semanticsand performance. How to provide enough file I/O performance and prove the correctness offile I/O operations for applications with different requirement have been the key issues ofnetwork filesystem design.With the design and implementation of Dawning cluster filesystem, this thesis deeplyexplored the basic principle to improve network filesystem performance, and prove thecorrect result of applications. The content of research includes the followings : research onthe influence of the communication subsystem to integral system performance and buildingCOSMOS network filesystem prototype, adjusting the network filesystem architecture andimproving the intensive file meta-data operations' performance of information servicesworkload, extending file lock interface to prove the correct execution of MPI-IO parallelcomputing applications. The detailed research results are the followings:1. This thesis firstly proposed the concept of concurrent bandwidth utilization. Using it,people can accurately evaluate the relationship between system efficiency and the networkcommunication system, I/O subsystem in servers. It can help people design and deploynetwork filesystem more reasonable.2. Give some approaches to improve the concurrent bandwidth utilization. We appliedBCL-3 and Myrinet in a prototype based on Dawning 3000 superserver, and studied andmeasured the influence of communication subsystem for concurrent bandwidth utilization.The performance results show fairly higher bandwidth are acquired for the filesystemclients comparing to prototype system based on 100Mbps Ethernet and TCP/IP, and theconcurrent bandwidth utilization is increased from less than 40% to about 90%.3. This thesis presented the FPLS and FPLS+ pathname lookup protocol to acceleratethe meta-data access by adjusting meta-data layout structure and improving the pathnamelookup algorithm in client node kernel. From the results of simulation, the new pathnamelookup protocol can shorten the time of pathname resolution to the 20.2 percent of the timenormal pathname resolution need.