Intelligent, Adaptive Parallel File Systems

点击翻译按钮获取中文摘要

The widening disparity between computation and input/output rates is now a major impediment to achieving high performance for a wide range of national challenge applications. This is especially true for emerging multidisciplinary applications with dynamic, irregular behavior and time varying input/output demands. An extensive analysis of application and physical input/output patterns by multiple research groups has shown that no single parallel file policy is likely to yield high performance for all input/output patterns. Instead, it will be necessary to aggressively exploit input/output access pattern knowledge in caching and prefetching systems to obtain a substantial fraction of peak hardware input/output performance. This project will explore the research issues underlying creation of (a) automatic behavioral classification techniques to identify and group application resource request patterns, (b) flexible file system policies that can trade storage space for increased input/output bandwidth and balance conflicting resource demands within and across applications, and (c) a set of fuzzy logic decision mechanisms for selecting file system policies based on application input/output requests and input/output system responses.