GPU fast multipole method with lambda-dynamics features

A significant and computationally most demanding part of molecular dynamics simulations is the calculation of long-range electrostatic interactions. Such interactions can be evaluated directly by the naïve pairwise summation algorithm, which is a ubiquitous showcase example for the compute power of graphics processing units (GPUS). However, the pairwise summation has O(N^2) computational complexity for N interacting particles; thus, an approximation method with a better scaling is required. Today, the prevalent method for such approximation in the field is particle mesh Ewald (PME). PME takes advantage of fast Fourier transforms (FFTS) to approximate the solution efficiently. However, as the underlying FFTS require all-to-all communication between ranks, PME runs into a communication bottleneck. Such communication overhead is negligible only for a moderate parallelization. With increased parallelization, as needed for high-performance applications, the usage of PME becomes unprofitable. Another PME drawback is its inability to perform constant pH simulations efficiently. In such simulations, the protonation states of a protein are allowed to change dynamically during the simulation. The description of this process requires a separate evaluation of the energies for each protonation state. This can not be calculated efficiently with PME as the algorithm requires a repeated FFT for each state, which leads to a linear overhead with respect to the number of states. For a fast approximation of pairwise Coulombic interactions, which does not suffer from PME drawbacks, the Fast Multipole Method (FMM) has been implemented and fully parallelized with CUDA. To assure the optimal FMM performance for diverse MD systems multiple parallelization strategies have been developed. The algorithm has been efficiently incorporated into GROMACS and subsequently tested to determine the optimal FMM parameter set for MD simulations. Finally, the FMM has been incorporated into GROMACS to allow for out-of-the-box electrostatic calculations. The performance of the single-GPU FMM implementation, tested in GROMACS 2019, achieves about a third of highly optimized CUDA PME performance when simulating systems with uniform particle distributions. However, the FMM is expected to outperform PME at high parallelization because the FMM global communication overhead is minimal compared to that of PME. Further, the FMM has been enhanced to provide the energies of an arbitrary number of titratable sites as needed in the constant-pH method. The extension is not fully optimized yet, but the first results show the strength of the FMM for constant pH simulations. For a relatively large system with half a million particles and more than a hundred titratable sites, a straightforward approach to compute alternative energies requires the repetition of a simulation for each state of the sites. The FMM calculates all energy terms only a factor 1.5 slower than a single simulation step. Further improvements of the GPU implementation are expected to yield even more speedup compared to the actual implementation.2021-11-1