SOFTWARE: Automatic Resource Selection in Dynamic Networked Computation Environments

该项目将对计算网格等动态共享网络环境中的资源选择中间件进行基础研究，该方法的特点是将应用资源需求与网络上可用的资源进行匹配，而不是基于固定的标准来选择最佳节点。关键挑战是估计给定应用程序在给定网络条件下在给定节点集上的运行情况。该项目将调查两种不同的方法来应对这一挑战。第一种是基于“异质性能模型”的模型，该模型预测不同资源可用性场景下的应用执行时间，例如不同速度的处理节点和网络链路以及不同的竞争负载和流量。第二种是基于“性能骨架程序”的概念，它是一个综合生成的短时间运行的程序，具有与给定应用程序相同的基本执行特征。在相同的网络条件下，性能骨架程序的执行时间预计是其所代表的应用程序预期执行时间的固定小部分。这两种方法的实现都基于对应用程序在执行期间使用的资源的实验测量。该过程基于监控在受控试验台上执行期间由应用程序产生的处理器、网络和存储器活动。该项目旨在为网格计算环境开发一个完整的资源选择框架，其基础是将这些应用信息模型与从网络气象服务等公开可用的网络测量和预测工具获得的网络状态信息进行匹配。

This project will perform fundamental research in middleware for resource selectionin dynamically shared network environments such as computation grids.The distinguishing feature of our approach is that it matches the application resourcerequirements to the resources available on a network rather than selecting the "best"nodes based on a fixed criterion. The key challenge is to estimate how a given application will perform on a given set of nodes under given network conditions. The project will investigate two different approaches to address this challenge. The first one is based on a "heterogeneous performance model", which is a model to predict application execution time under different resource availability scenarios, such asdifferent speed processing nodes and network links and different competing loads andtraffic. The second is based on the concept of a "performance skeleton program",which is a synthetically generated short running program that has the same fundamental execution characteristics as a given application. Under the same network conditions, the execution time of a performance skeleton program is expected to be a fixed small fraction of the expected execution time of the application it represents. The implementation of both these approaches is based on experimental measurement of the resources that an application uses during execution. The procedure is based on monitoring the processor, network and memory activity generated by an application during execution on a controlled testbed. The project aims to develop a complete resource selection framework for grid computing environments based on matching these application information models to the network status information obtained from publicly available network measurement and prediction tools like the Network Weather Service.