NGS: Grid Harvest Service (GHS): A Performance System for Grid Computing

网格计算带来了地理上分散的计算能力，以满足日益增长的需求。然而，代表着新的和未来的计算平台，网格是一个全球共享的、异构的、自主控制的平台。它比过去的计算平台更先进、更强大、更动态、更复杂。这种复杂性需要新的系统软件技术来有效地利用其计算能力。大多数现有的性能技术都是针对专用平台的。最近的性能工具，如NWS，只预测短期(不到五分钟)的资源可用性，这不适合长期应用程序。需要新的软件技术来进行长期的、应用级的性能预测和任务调度，以减轻网格的复杂性。初步结果表明，GHS从根本上来说比现有的系统更适合长期应用，并可以大幅降低计算成本。PIS将与能源部国家实验室和IIT的研究人员合作，展示GHS具有重要的国家利益应用的巨大潜力。研究方法是基于这样的观察，即网格环境没有中央控制，性能效率必须基于资源的可用性。GHS将利用这一观察结果。具体来说，该项目：1)将设计、实施和验证随机和分析模型，以预测计算和通信资源的可用性及其对用户应用程序的影响。2)将开发、实施和验证实用的、非侵入性的绩效测量技术。3)将设计、实现和验证任务调度和重新调度算法，以利用(1)中给出的预测来减少用户应用程序运行时间。

Grid computing brings geographically dispersed computing power to meet the ever-increasing demands. However, representing new and future computing platforms, Grid is a globally shared, heterogeneous, and autonomous controlled platform. It is far more advanced, powerful, dynamic and complex than computing platforms in he past. Such complexity requires new system software technology to efficiently utilize its computing capacity. Most existing performance technologies are targeted for dedicated platforms. Recent performance facilities, such as NWS, only predict short-term (less than five minutes) resource availability, which is not appropriate for long-term applications. New software technologies are needed for long-term, application-level performance predication and task scheduling to alleviate the complexity of Grid. Preliminary results show that GHS is fundamentally better than existing systems for long-term applications and can lead to substantial decrease in computing cost. The PIs will collaborate with researchers at DOE national laboratories and IIT to demonstrate the great potential of GHS with important national interest applications. The research approach is based on the observation that Grid environments do not have central control and performance efficiency has to be based on resource availability. GHS will exploit this observation. In particular the project: 1) will design, implement, and validate stochastic and analytical models to predict the computing and communication resource availability and their influence on user applications. 2) will develop, implement, and validate practical and non-intrusive performance measurement technologies. 3) will design, implement, and validate task scheduling and rescheduling algorithms to utilize the prediction given in (1) to reduce user application run-times.