Development of high performance computation system with great precision for molecular dynamics simulations

分子动力学模拟为研究分子体系的结构和动力学提供了有价值的手段。由于它需要大量的计算资源，特别是当它应用于大型生物系统时，它给计算技术带来了具有挑战性的问题。虽然已经开发了几种快速算法来减少计算时间，但这些算法都遇到了一定的困难。在这项工作中，我首先重新制定了快速多极子方法(FMM)，以进一步减少计算时间和内存需求。其次，在FMM框架下，提出了两种可靠的周期性算法。一种方法使用类似Ewald的技术，另一种方法使用宏观的细胞层次结构。由于包括FMM在内的每一种快速算法都使用某种近似，因此算法在某些情况下会出现不准确的情况。在这项工作中，我发现对FMM中涉及的任务使用专用机器可以提高精度和计算速度。这样一个算法新颖的计算系统正是在这项工作中开发出来的。经证实，该系统的性能比普通工作站高出几百倍。

Molecular dynamics simulation provides valuable means to study the structure and dynamics of molecular systems. Since it demands vast computational resources, particularly when it is applied to large biological systems, it presents challenging problems to computational technology. Although several fast algorithms were developed for reducing the computation time, these algorithms encounter certain difficulties. In this work, I first reformulated the fast multipole method (FMM) for further decreasing the computation time and memory requirements. Secondly, I developed two reliable algorithms for accomplishing periodicity in the FMM framework. One method employs the Ewald-like technique, and the other uses a macroscopic hierarchical structure of cells. These methods developed in this work enable reliable simulations of biological macromolecules.Since every fast algorithm including FMM employs a certain kind of approximation, the algorithms suffer from inaccuracy in some situation. In this work, I found that use of a special-purpose machine for a task involved in FMM improve both accuracy and computation speed. Such a computation system with novel algorithms was actually developed in this work. The performance of the system was confirmed to be several hundred times higher than those of ordinary workstations.