High Performance Algorithms for Electronic Materials

电子材料的高性能算法

• 批准号: 9525885
• 负责人: James Chelikowsky
• 金额: $ 59.3万
• 依托单位: University of Minnesota-Twin Cities
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 1995
• 资助国家: 美国
• 起止时间: 1995-12-01 至 1999-11-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=9525885&HistoricalAwards=false
• 关键词: Performance; Algorithms; Electronic; Materials

9525885 Chelikowsky This computational materials science grant is co-funded by the Materials Theory Program in the Division of Materials Research and the New Technologies Program in the Advanced Scientific Computing Division. One principal investigator is a computational materials scientist while the other is a computational scientist. A primary objective of this research is to develop computer codes for materials modeling on advanced platforms which utilize the combined expertise of these two researchers. The goal of the research is to develop computational methods which fully utilize innovations in high performance computers in order to open new applications of electronic structure codes to complex materials. The focus of the research will be on real space methods through higher order finite difference techniques to electronic structure problems. Finite difference techniques coupled with ab initio pseudopotentials can provide a simple, yet accurate and efficacious, approach to complex systems. A few of the advantages of finite difference techniques over more traditional plane wave techniques include: finite difference techniques are far easier to implement than plane wave codes with no loss of accuracy, especially for parallel implementations; these techniques to not require the use of supercells for localized systems. No cell-cell interactions are present. Charged systems can be handled directly without artificial compensating backgrounds. Replication of vacuum is natural and minimized compared to extended basis sets; finite difference algorithms result in a minimum factor of 4-5 gain in speed over plane wave techniques for shared memory architectures. Significantly better performance is expected for massively parallel platforms; no fast-Fourier transforms are required. Global communications are minimized. Applications will center on electronic materials such as large clusters, liquids and amorphous s olids. Algorithm development will concentrate on designing parallel preconditioners for the eigenvalue problem, as well as the implementation of these techniques on various paralllel platforms. %%% This computational materials science grant is co-funded by the Materials Theory Program in the Division of Materials Research and the New Technologies Program in the Advanced Scientific Computing Division. One principal investigator is a computational materials scientist while the other is a computational scientist. A primary objective of this research is to develop computer codes for materials modeling on advanced platforms which utilize the combined expertise of these two researchers. The goal of the research is to develop computational methods which fully utilize innovations in high performance computers in order to open new applications of electronic structure codes to complex materials. ***

本计算材料科学基金由材料研究部材料理论项目和高级科学计算部新技术项目共同资助。一位首席研究员是计算材料科学家，另一位是计算科学家。本研究的主要目标是利用这两位研究人员的综合专业知识，为先进平台上的材料建模开发计算机代码。研究的目标是开发充分利用高性能计算机创新的计算方法，以打开复杂材料的电子结构代码的新应用。研究的重点将是利用高阶有限差分技术解决电子结构问题的实空间方法。有限差分技术与从头算伪势相结合，可以为复杂系统提供一种简单、准确和有效的方法。有限差分技术相对于传统的平面波技术的一些优点包括：有限差分技术比平面波代码更容易实现，而且没有精度损失，特别是对于并行实现；这些技术不需要在局部系统中使用超级细胞。没有细胞间的相互作用。带电系统可以直接处理，而无需人工补偿背景。与扩展基集相比，真空的复制是自然的和最小化的；有限差分算法导致共享内存架构的速度比平面波技术增加4-5倍的最小因子。大规模并行平台的性能有望显著提高；不需要快速傅里叶变换。全球通信被最小化。应用将集中在电子材料，如大簇，液体和无定形固体。算法开发将集中于设计特征值问题的并行预调节器，以及这些技术在各种并行平台上的实现。这项计算材料科学基金由材料研究部的材料理论项目和高级科学计算部的新技术项目共同资助。一位首席研究员是计算材料科学家，另一位是计算科学家。本研究的主要目标是利用这两位研究人员的综合专业知识，为先进平台上的材料建模开发计算机代码。研究的目标是开发充分利用高性能计算机创新的计算方法，以打开复杂材料的电子结构代码的新应用。＊＊＊