ALGORITHMS: Hierarchical Computational-space Decomposition: A Framework for Scalable Scientific Computing Beyond Teraflop

算法：分层计算空间分解：超越 Teraflop 的可扩展科学计算框架

• 批准号: 0203363
• 负责人: Aiichiro Nakano
• 金额: $ 20.5万
• 依托单位: Louisiana State University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2002
• 资助国家: 美国
• 起止时间: 2002-09-01 至 2002-10-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0203363&HistoricalAwards=false
• 关键词: ALGORITHMS; Hierarchical; Computational; space; Decomposition

This project will develop a scalable parallel computing framework for high-end computational research, which will achieve scalability beyond tightly coupled Teraflop architectures, i.e., for distributed supercomputing on multiple Teraflop computers as well as on future Petaflop computers with deep memory hierarchy. To accomplish this goal, the PI will conduct the following research tasks: Topology-preserving computational-space decomposition to minimize the number of messages using a structured message-passing scheme; Wavelet-based adaptive load balancing in dynamic, heterogeneous metacomputing environment using simulated annealing to minimize load imbalance and message sizes; Recursive and reconfigurable grouping of processors with message renormalization and computation/communication overlapping to hide latency at each grouping level; Spacefilling-curve-based adaptive data compression --- in situ processing of interoperable compressed data to reduce message sizes with user-controlled error bound. A suite of scalable scientific programs developed within the new framework will be disseminated as a computational-scientist's toolkit through a Web portal to have significant impacts on high-end computational research, including the design of quantum-dot architectures for future quantum computing.

该项目将为高端计算研究开发一个可扩展的并行计算框架，该框架将超越紧耦合的teraflop体系结构，即在多teraflop计算机上以及在未来具有深层存储层次的petaflop计算机上实现分布式超级计算。为了实现这一目标，PI将进行以下研究任务：拓扑保持计算空间分解，以使用结构化的消息传递方案来最小化消息的数量；基于小波的自适应负载平衡，在动态、异构的元计算环境中，使用模拟退火法来最小化负载失衡和消息大小；递归和可重新配置的处理器分组，消息重整化和计算/通信重叠以隐藏每个分组级别的延迟；基于空间填充曲线的自适应数据压缩--原位处理可互操作的压缩数据，以减小消息大小，并具有用户控制的误差界。在新框架内开发的一套可扩展的科学程序将作为计算科学家的工具包通过Web门户网站传播，对高端计算研究产生重大影响，包括为未来量子计算设计量子点架构。