CAREER: of Computational and Data Access Scheduling in NOWs

工作站网络(NOW)已成为科学计算的主要基础设施。NOW目前用于提供桌面顺序计算和可视化、批量顺序作业的高吞吐量机会计算以及专用并行计算。此外，我们设想NOW将用于使用科学数据库管理系统(DBMS)存储和检索数据集，提供并行程序的机会性计算，允许对大型磁盘驻留数据集的交互可视化。目前的知识状态提供了实现这种设想的Now使用的大多数工具，但缺乏所需的调度策略。为了解决这个问题，该项目将：*设计新的调度算法来处理混合工作负载，包括工作站作业、批处理机会性顺序作业和批处理机会性并行作业。*描述科学数据库管理系统的工作负载，并就如何最好地将科学数据库管理系统整合到现在进行设计。*确定如何最好地整合索引技术，以支持大型磁盘驻留数据集的交互可视化。*制定系统范围的调度策略，了解NOW使用的所有方面。此类策略的目标是为所有并发运行的应用程序提供良好的聚合性能，包括计算、可视化和数据访问。拟议的工作将开发包括上述所有因素在内的NOW性能的全面分析和模拟模型，以确定如何优化整体性能，并将开发该系统的工作原型。

Networks of workstations (NOWs) have been emerged as a primary infrastructure for scientific computing. NOWS are currently used to provide desktop sequential computing and visualization, high throughput opportunistic computing of batched sequential jobs, and dedicated parallel computing. In addition, we envision NOWs being used to store and retrieve data sets using scientific database management systems (DBMSs), provide opportunistic computation of parallel programs, allow interactive visualization of large disk resident data sets. The current state of knowledge provides most of the tools to enable this envisioned NOW usage, but lacks the needed scheduling policies. To remedy this, this project will: * Devise new scheduling algorithms to handle mixed workloads of workstations jobs, batched opportunistic sequential jobs, and batched opportunistic parallel jobs. * Characterize scientific DBMS workloads and develop designs on how best to integrate scientific DBMSs into the NOW. * Determine how best to integrate indexing techniques to provide support for interactive visualization of large disk resident data sets. * Develop system wide scheduling policies that are cognizant of all aspects of NOW usage. The goal of such policies is to provide good aggregate performance of all concurrently running applications including computation, visualization, and data access. The proposed work will develop comprehensive analytical and simulation models of NOW performance, including all of the above factors, to determine how to optimize overall performance and will develop a working prototype of the system.