Postdoc: Exposing Input/Output Parallelism by Relaxing Request Ordering

97-04704加思·吉布森卡内基梅隆大学博士后：通过放松请求排序来暴露输入/输出并行性对于许多并行科学应用程序来说，输入/输出是一个关键的性能限制。从应用程序性能分析中学到的一个重要教训是，公开和利用所有可用的输入/输出并发性非常重要。传统的顺序文件模型不允许应用程序表达其资源需求的并行性，也不允许充分利用并行文件系统带宽。这项研究提出了新的应用级输入/输出接口来表达无法自动检测的并行性。将在各种并行平台上创建和评估访问具有部分有序或无序记录的并行文件的接口，量化文件系统内各个级别实现的性能改进。应用程序所展示的好处提供了访问模式信息；而文件系统策略控制驱动程序试图标准化系统和应用程序级编程接口。然而，人们对系统级文件系统功能和应用程序级文件系统功能之间的划分知之甚少。这项研究有助于为一大类重要的并行科学应用程序定义这一边界。

97-04704 Gibson, Garth A Carnegie Mellon University Postdoc: Exposing Input/Output Parallelism by Relaxing Request Ordering (CES Associateship for Tara Madhyastha) Input/output is a critical performance limitation for many parallel scientific applications. One important lesson learned from application performance analysis is that it is important to expose and utilize all available input/output concurrency. A traditional sequential file model does not allow applications to express the parallel nature of their resource demands, or to exploit the full parallel file system bandwidth. This research proposes new application-level input/output interfaces to express parallelism that cannot be automatically detected. Interfaces to access parallel files with partially ordered or unordered records will be created and evaluated on a variety of parallel platforms, quantifying the performance improvements achieved at various levels within the file system. The demonstrated benefits of application supplies access pattern information; while file system policy control drive attempts to standardize system and application-level programming interfaces. However, the division between system and application-level file system functionality is poorly understood. This research helps to define this boundary for a large and important class of parallel scientific applications.