Biophysical Society Thematic Meeting | Hamburg 2022

Biophysics at the Dawn of Exascale Computers

Poster Abstracts

27-POS Board 27 FAST MULTIPOLE METHOD FOR A CONSTANT PH ALGORITHM IN GROMACS Bartosz Kohnke 1 ; Eliane Briand 1 ; Carsten Kutzner 1 ; Helmut Grubmüller 1 ; 1 Max Planck Institute for Multidisciplinary Sciences, Theoretical & Computational Biophysics, Göttingen, Germany Molecular dynamics (MD) simulations of biomolecules with dynamic protonation changes have lately become increasingly important. Constant pH molecular dynamics (CPHMD) allows to dynamically alter the protonation during simulations to correctly model protonation probabilities at a fixed pH level. Typically, CPHMD utilizes a λ -dynamics method with Hamiltonian scaling, where a fictitious λ -particle continuously interpolates between protonated and deprotonated Hamiltonian. The dynamics of the λ -particle depends on the local electrostatic environment of a protonatable site; however, the long-range nature of coulombic forces requires reevaluation of the complete Hamiltonian at each protonation state. This step becomes a prohibitive factor for efficient usage of CPHMD systems with many sites since pairwise interactions evaluation is the most performance-limiting factor in MD simulations. To obtain atomistic trajectories of proteins on biologically relevant timescales, MD utilizes efficient approximation algorithms. The most prominent one in the field is particle mesh Ewald (PME), which is extremely fast but lacks the flexibility to perform Hamiltonian recalculations efficiently. A slightly different approach to CPHMD is a charge scaling method, which allows a more efficient interpolation scheme and a proper PME utilization. However, to enable efficient λ -dynamics with Hamiltonian scaling, we implemented a fast multipole method (FMM). In contrast to PME, FMM utilizes a flexible tree structure that preserves local charge differences, and therefore it allows for very fast Hamiltonian recalculations. Here, we present our NVIDIA CUDA FMM for CPHMD. The FMM GPU implementation has been tailored for MD calculations. Its performance is about a third of that of highly optimized PME on a single GPU node. However, the FMM allows simulation systems with many protonatable sites with negligible computational overhead. Additionally, on larger exascale GPU clusters, where PME scaling breaks down, the efficiency of FMM for CPHMD should become even more apparent.

96