Application of lightweight threading techniques to computational chemistry

Abstract

The recent advent of inexpensive commodity multiprocessor computers with standardized operating system support for lightweight threads provides computational chemists and other scientists with an exciting opportunity to develop sophisticated new approaches to materials simulation. We contrast the flexible performance characteristics of lightweight threading with the restrictions of traditional scientific supercomputing, based on our experiences with multithreaded molecular dynamics simulation. Motivated by the results of our molecular dynamics experiments, we propose an approach to multi-scale materials simulation using highly dynamic thread creation and synchronization within and between concurrent simulations at many different scales. This approach will enable extremely realistic simulations, with computing resources dynamically directed to areas where they are needed. Multi-scale simulations of this kind require large amounts of processing power, but are too sophisticated to be expressed using traditional supercomputing programming models. As a result, we have developed a high-level programming system called Sthreads that allows highly dynamic, nested multithreaded algorithms to be expressed. Program development is simplified through the use of innovative synchronization operations that allow multithreaded programs to be tested and debugged using standard sequential methods and tools. For this reason, Sthreads is very well suited to the complex multi-scale simulation applications that we are developing.