to the one taken in MD. In particular, they describe the linked-cell method of distributing computation in molecular dynamics simulations. As in MD, the linkedcell method assigns particles to processors based on their location in three-dimensional cells (domains in MD) of size greater or equal to cutoff distance. In addition to the exploitation of molecular locality, the algorithm presented here differs from the linked-cell approach in that it supports domains of arbitrary shapes and has no restrictions on the sizes of those domains.
MD is written using Cthreads, a multiprocessor threads package that supports thread-based parallelism and provides portability between a variety of shared-memory multiprocessor and uniprocessor platforms. Although MD is currently implemented on shared-memory multiprocessors, its basic structure is easily adapted for and ported to distributed-memory machines. In addition to this portability, MD offers substantial flexibility in its support for experimentation with alternative domain decompositions or synchronization methods, or with alternative approaches to exploiting molecular locality. Accordingly, an important contribution of this paper is the presentation of various aspects of MD's design, the tradeoffs involved and their effects on performance. These discussions form the bulk of Sections 3 and 4 and serve to support the main conclusions of this work.
One important conclusion of our research is that molecular locality can be successfully exploited to reduce computational complexity. A second conclusion is that flexibility in terms of being able to support multiple elements, molecular systems, schemes for domain decomposition, and communication/synchronization methods need not increase computational costs. On the contrary, such flexibility is important in order to attain high performance for applications written with MD.
Examining the design and behavior of MD is useful for several reasons. First, the application itself is of interest to molecular dynamicists because many similar molecular dynamics programs can be built within the modular framework of MD. Second, applications designers may benefit from an understanding of the tradeoffs made in MD's implementation. Third, MD provides computer scientists with a portable, complex, and dynamic sample high performance application for use in the investigation of topics in operating or programming systems for high performance parallel machines.