Next Generation Software: Coordinated Allocation of Processor and I/O Resources in Parallel Systems

点击翻译按钮获取中文摘要

EIA-9974992Smirni, EvgeniaCollege of William & MaryNext Generation Software: Coordinated Allocation of Processor and I/O Resources in Parallel SystemsA problem that plagues most parallel systems today is inefficient management of resources. Parallel systems are frequently over-utilized, yet they deliver relatively low performance to individual applications. In recent years, much research had been devoted to methods for allocating the processors of a parallel system to competing applications. However, an issue that has been traditionally overlooked is that many applications also contend for shared Input/Output (I/O) resources. Since I/O resources belong to the slowest level of memory hierarchy, their efficient management becomes critical for high performance.The proposed research focuses on the development of an allocation framework that integrates the management of computational and I/O resources in parallel systems. This research is facilitated, firstly, by previously obtained quantitative data on the processing and I/O requirements of parallel scientific applications. Additionally, recent studies on processor scheduling that explicitly consider the I/O demands of parallel scientific applications demonstrate performance trends that differ significantly from previously observed trends where only processor allocation is considered.The proposed plan considers a systematic exploration of coordinated allocation strategies of both processor and I/O resources and investigates issues that have not been addressed in traditional processor scheduling or I/O research in parallel systems. The primary objectives of this project are: 1) the development of an analytic model that effectively captures the scalability of I/O intensive applications under different data distributions on the disks and under different assignment of computational resources, 2) the analysis of the relative advantages and disadvantages of new dual resource allocation strategies through modeling and experimental measurements, and 3) the development of a unified framework that allows the processor scheduler and the parallel file system to operate jointly towards the common goal of sharing the parallel resources amoung the various executing applications in order to maximize overall system performance. The proposed work will be experimentally verified on a variety of parallel platforms and the dual resource allocation algorithms will be implemented so as to ensure their portability across these platforms.