High performance I/O architectures and file systems for Internet servers

Previous studies have shown that disk I/O is one of the major performance bottlenecks of Web proxy servers, one class of Internet servers. Most conventional file systems and I/O architectures do not work well because of high overheads. This dissertation makes several contributions to solve this problem.; First, it studies Log-structured File System (LFS) as one possible existing solution. We find two ways to significantly improve LFS performance: one is a novel reordering write buffer called WOLF, which is being developed to reduce LFS garbage collection overheads by classifying data into two classes in the write buffer and forcing a bimodal distribution on the disk; the other is a data reorganization policy called PROFS, which integrates disk physical geometries to effectively optimize data placements on multiple zone disks.; Second, we develop a novel, U&barbelow;ser-space, C&barbelow;ustomized F&barbelow;ile S&barbelow;ystem called UCFS that can boost I/O performance for Web proxy servers. UCFS is a user-level software component of a proxy server which manages data on a raw disk or disk partition. It is not only easy and inexpensive to be implemented, but also has good portability and maintainability. UCFS uses several novel designs, including a novel file system called Cluster-structured File System (CFS) to provide persistent high I/O performance without degradation. Comprehensive experiments show that UCFS can markedly improve I/O performance for Web proxy servers by more than one order of magnitude.; In the future, WOLF and PROFS can be successfully applied to those I/O storage systems that use log-structuring techniques while UCFS would be prosperously extended to other areas such as Web servers and mail/news servers.