Developing a new molecular dynamics code with an eye towards portability

Classical molecular dynamics has become a ubiquitous computational modeling tool for a number of disciplines, from biology and biochemistry, to geochemistry and polymer physics. Due to intense efforts from a number of developers over the past 50 years, several MD programs have been highly successful in achieving commendable efficiency and overall performance.

The classical molecular dynamics algorithm involves three main components: the integration step, the calculation of short-range forces, and the calculation of long-range forces. The integration step is generally the quickest part of the calculation, and as it has some memory-intensive aspects, is often calculated using the CPU, in implementations using heterogeneous architectures. The long-range force calculation, in most implementations, involves an Ewald sum. This requires the use of Fourier transform methods, which are fast for smaller systems, but do not scale well for large systems. This is an active area of development and is not addressed here. The major bottleneck for all system sizes is the short-range non-bonded forces (SNFs) calculation, as it involves a sum of pairwise interactions over multiple subsets of the particle space.

As part of our portable performance studies, we have written a new SNF kernel, wherein we use directives (OpenACC) to implement the parallel steps of the computation. We have also produced an alternate implementation where matrix-matrix multiplication is used to calculate pairwise distances in the SNF calculation. This alternate implementation, though requiring more floating-point operations, is shown to perform well because of the performance of platform-specific BLAS libraries.

Details of the MD experiment can be found in this report.