The Design And Implementation Of The Parallel Network File System PNFS

The data in large-scale parallel computing has increased with volatile speed. It forces the I/O storage technology to develop from Direct-Attached Storage (DAS) to Network-Attached Storage (NAS) and Storage Area Network (SAN). Although SAN has high performance, it can not meet the requirements of the high performance computing (HPC) for its low cost-effective. With the development of IP network, especially for Ethernet technology, NAS is becoming more cost-effective and the network bandwidth develops quickly from 10Mbps to 1000Mbps. Now, 10000Mbps technology has appeared and applied in some fileds. However, the application of NAS is limited by the centralized architecture of NFS. Therefore, it is an urgent demand to look for a solution to organize the current NAS storage equipments effectively. Besides the original advantages of NAS, this solution will get higher performance, higher scalability, higher safety and higher manageability to satisfy the demands of HPC.In this thesis, the architecture of parallel network file system (PNFS) is designed based on the analyses and researches of various parallel file systems and their key techniques. PNFS enables the separation of the metadata server and the storage server to eliminate the bottleneck caused by the centralized architecture. PNFS is the extension of NFSv4 protocol, and provides the ability to access file data parallel. PNFS can get the linear increases on the I/O bandwidth with the increasing number of storage servers. PNFS is composed of three parts: PNFS client, PNFS metadata server and storage server. In this thesis, the design and implementation of the three parts are detailed in the following chapters.Finally, this thesis gives the performance tests of PNFS, which is made on the LAN environment of 1000Mbps Ethernet. The results show that PNFS can provide a relatively high I/O read/write bandwidth with low latency, and can meet the I/O requirements of HPC.