Halide in Molecular Dynamics

Abstract

In many fields, especially biology and chemistry, it is important to understand how a collection of particles will interact with each other over some period of time. If only managing a system of two particles, it is simple enough to calculate the final positions of the atoms given their properties and the outside forces placed upon them. However, it is often the case that the system’s size is several magnitudes larger; therefore, the task is handed off to computers and simulators.

Molecular dynamics, or MD, simulations tend to be extremely expensive, taking several weeks to compute less than a second’s worth of real time. Two significant reasons MD simulations are time intensive are due to the complex loop structures and math required to observe each time step. More tools and research are constantly being developed to increase performance of these simulations.

In this thesis we introduce a tool from the image processing domain, Halide, and argue that Halide is a qualified candidate to efficiently implement MD simulations in the future. We rewrote a potential into Halide to achieve only a 20% slow down serially, which we are certain can reach parity with minimal changes to the code, and over 300% speed up when running in parallel. While it was challenging beginning to work with Halide and its limitations, we are still able to accomplish this performance and versatility writing 47% less code. Halide also makes the transformation to parallel scheduling trivial, whereas this is not the case in the original implementation. Halide was not able to represent all of the loop structures we wanted; however, we also suggest several additions and changes to Halide to make it more suitable to MD.