CHARMM Modernization, Performance, and Continued Development

描述(申请人提供)：二十多年来，CHARMM(哈佛大学大分子力学化学)一直是生物学中大分子模拟和建模的主要研究工具。在此期间，CHARMM的发展，以及从其使用中出现的应用，定义了生物分子计算领域。这项提议旨在确保CHARMM将继续作为未来几代科学家的开发平台，方法是解决软件组件及其附带的支持基础设施，这些组件是开发、性能和维护方面的瓶颈。具体地说，重组代码将改进CHARMM中的底层计算内核，这将显著增强单处理器性能并改善并行伸缩性。重点是提高两个目标系统规模的并行度。对于代表许多“典型”应用的50-200K原子系统，我们将开发、调整和部署在“部门可访问”并行平台(约100个具有高带宽互连的商用处理器)上实现分子动力学紧耦合并行的良好并行效率(大于70%)的方法。对于大型“超级计算机”上的大型生物学问题，即运行在0.5-2K处理器上的接近1M个原子的系统，我们将开发和实现新的CHARMM模拟内核，利用这些系统的大空间维度，并使用空间分解和任务级并行技术。最后，我们将开发和改进代码和算法，允许基于图形处理器的加速用于CHARMM潜力和方法家族。这些努力的结果将是一个程序平台，它将促进在大分子模拟和建模方面的持续前沿研究，并使其能够为未来几代研究人员继续开发和维护。

DESCRIPTION (provided by applicant): CHARMM (Chemistry at Harvard Macromolecular Mechanics) has been a primary research tool for macromolecular simulations and modeling in biology for over two decades. During this period CHARMM development, and applications emerging from its use, have defined the field of biomolecular computation. This proposal is aimed at ensuring CHARMM will continue as the development platform for future generations of scientists by addressing components of the software and its attendant support infrastructure that represent bottlenecks in development, performance and maintenance. Specifically, restructuring the code will provide improvements to underlying computational kernels in CHARMM, which will significantly enhance single processor performance and improve parallel scaling. Emphasis will be given to improving parallelism for two target system sizes. For systems of 50-200K atoms, which represent many "typical" applications, we will develop, adapt and deploy methods that yield good parallel efficiency (greater than 70%) for tightly coupled parallelism in molecular dynamics on "department accessible" parallel platforms (~100 commodity processors with a high bandwidth interconnect). For large biological problems on large "supercomputers", i.e. for systems approaching 1M atoms running on 0.5-2K processors, we will develop and implement new simulation kernels for CHARMM that exploit the large spatial dimensions of these systems and employ techniques of spatial decomposition and task-level parallelism. Finally, we will develop and improve code and algorithms allowing graphics processor based acceleration to be utilized for the family of CHARMM potentials and methods. The outcome of these efforts will be a program platform that will facilitate continued forefront research in macromolecular simulation and modeling and enable its continued development and maintenance for future generations of researchers.