<span class="var-sub_title">PotC: Many-Body Potential Implementations à La Carte</span> SC18 Proceedings

The International Conference for High Performance Computing, Networking, Storage, and Analysis

PotC: Many-Body Potential Implementations à La Carte

Student: Markus Höhnerbach (RWTH Aachen University)
Supervisor: Paolo Bientinesi (RWTH Aachen University)

Abstract: Molecular dynamics is a valuable investigation tool for simulations in computational chemistry and materials science. In these simulations, atoms move according to so-called potentials, functions that typically describe the distance-dependent interactions between pairs of atoms. For some applications more complex "many-body potentials" are required. From a HPC perspective, such potentials pose challenges: small neighborhoods hinder vectorization; redundant force expressions are tedious and error prone to derive; the implementations are large, runtime-critical, and can not be broken into simpler "kernels''. Consequently, only selected many-body potentials receive optimizations---with major effort.

PotC is a DSL for MD potentials and corresponding compiler to generate high-performance implementations. The compiler centralizes optimization knowledge for many-body potentials, and unburdens users from manual optimization and force derivation. The performance optimization of the generated code is still work-in-progress. However, independent of performance, the tool can also be applied when prototyping, testing, or possibly symbolically manipulating potentials.

ACM-SRC Semi-Finalist: no

Poster: PDF
Poster Summary: pdf
Reproducibility Description Appendix: PDF

Back to Poster Archive Listing