Realistic Simulations of Photoactive Systems on HPC Clusters with Many-Core Processors

这一跨学科项目旨在通过在下一代处理器平台上进行精确但计算要求高的模拟，来解决在理解光活性分子系统中能量转换的原子机制方面的挑战。最近关于光合作用复合体的实验表明，在更大的叶绿素网络中，支持从天线到反应中心的单向能量转移的部分相干机制。为了阐明其潜在的物理机制，并可能将其应用于有机太阳能电池等高效器件的设计，需要基于原子第一性原理计算的精确模拟。虽然在硫磺细菌中发现的最原始的光合作用复合体之一的时间分辨光谱的模拟可以在具有分层运动方程(HEOM)的单个图形处理器上进行，但更大的光活性分子系统需要具有多核处理器的高性能集群。这就需要新的策略来优化和扩展应用，以回答有关激子的时间尺度和选择的路径的基本问题。由于进一步提高单核性能的潜力有限，大规模并行多核体系结构正日益成为高性能计算集群的主力。从计算机科学的角度来看，模块化HEOM代码是研究开发当前和即将到来的多核系统的固有并行性的普遍适用方法的完美候选者。该项目将确定跨不同平台的通用优化策略，并将使用图论方法为共享和分布式内存布局提供避免通信的设计原则。

This interdisciplinary project aims to tackle challenges in the understanding of atomistic mechanisms of energy conversion in photoactive molecular systems through accurate but computationally demanding simulations on next generation processor platforms. Recent experiments on photosynthetic complexes suggest a partly coherent mechanism supporting unidirectional energy-transfer from the antenna to the reaction center in larger chlorophyll networks. To elucidate the underlying physical mechanism and to possibly apply it to the design of efficient devices such as organic solar cells, requires accurate simulations based on atomistic first-principle calculations.While the simulation of time-resolved spectra for one of the most primitive photosynthetic complexes found in sulfur bacteria can be performed on single graphics processors with the "Hierarchical Equations Of Motion" (HEOM),larger photoactive molecular systems require high-performance clusters with many-core processors. This demands for new strategies for optimizing and scaling the application to answer fundamental questions about the time-scale and chosen pathway of the excitons. Because of the limited potential of further improving the single-core performance, massively parallel many-core architectures are increasingly becoming the work-horse of high-performance compute clusters. From the computer science perspective, the modular HEOM code is a perfect candidate for studying generally applicable approaches to exploit the inherent parallelism of current and forthcoming many-core systems. The project will identify generic optimization strategies across heterogeneous platforms, and will contribute communication-avoiding design principles for shared and distributed memory layouts using graph theoretical methods.