Research Initiation Award: Job Scheduling and Memory Management in Message-passing Multicomputers

这个项目的主要目标是研究在多处理器环境中基于多编程的任务调度。这是基于这样一种信念，即像单处理器一样，多计算机必须进行多编程，以实现性能与成本的比。在多台计算机中多个应用程序之间共享资源的现有策略是静态的。这个项目将研究在设计和实现比静态策略性能更好的替代作业调度策略时出现的问题。作为第一步，将通过实验测量和分析来描述多计算机上的工作量。从本研究中获得的见解将用于自适应和动态分区策略的设计和实现。特别地，将研究使自适应和动态分区对应用程序程序员透明所涉及的实现问题。作为研究的一部分，将实现一个支持面向应用程序的动态重新配置的运行时系统。本研究的另一个目标是检查在较新的多计算机系统中引入虚拟内存对各种调度策略之间的性能权衡的影响。通过仿真和分析，确定几种科学应用的工作集大小和内存参考模式。该信息将用于基于模拟的研究，该研究将检查在系统的每个节点上支持虚拟内存的多台计算机中处理器分配、节点调度和内存管理之间的相互作用。*＊*

The main goal of this project is to investigate multiprogramming based task scheduling in multiprocessor environments. This is based on the belief that, like uniprocessors, multicomputers have to be multiprogrammed in order to achieve a performance to cost ratio. Existing policies for resource sharing in a multicomputer among multiple applications are static. This project will examine issues that arise in the design and implementation of alternate job scheduling policies that have a better performance than static policies. As a first step, the workload on multicomputers will be characterized through experimental measurement and analysis. The insights gained from this study will be used in the design and implementation of adaptive and dynamic partitioning policies. In particular, implementation issues involved in making adaptive and dynamic partitioning transparent to the applications programmer will be examined. As a part of the research, a run time system that provides support for application-directed dynamic reconfiguration will be implemented. Another goal of the research is to examine the impact of the introduction of virtual memory in newer multicomputer systems on the performance trade-offs between various scheduling policies. The working set sizes and memory reference patterns of several scientific applications will be determined through simulation and analysis. This information will be used in a simulation based study that examines the interactions between processor allocation, node scheduling and memory management in multicomputers that support virtual memory on each node of the system. ***