When studying membrane-bound protein receptors, it is necessary to move beyond the current state-of-theart simulations that only consider small membrane patches and implicit solvent. Limits of traditional computer platforms negatively impact the model’s level of realism and the computational scales achievable. On the other hand, multi-core platforms such as GPUs offer the possibility to span length scales in membrane simulations much larger and with higher resolutions than before.
To this end, this paper presents the design and implementation of an advanced GPU algorithm for Molecular Dynamics (MD) simulations of large membrane regions in the NVT, NVE, and NPT ensembles using explicit solvent and Particle Mesh Ewald (PME) method for treating the conditionally convergent electrostatic component of the classical force field. A key component of our algorithm is the redesign of the traditional PME method to better fit on the multithreading GPU architecture. This has been considered a fundamentally hard problem in the molecular dynamics community working on massively multithreaded architecture. Our algorithm is integrated in the code FENZI (yun dong de FEN ZI in Mandarin or moving molecules in English). The paper analyzes both the performance and accuracy of large-scale GPU-enabled simulations of membranes using FENZI, showing how our code can enable multi-nanosecond MD simulations per day, even when using PME.
In this paper we presented the design and implementation of FENZI, a MD code for the simulation of large membrane regions in both NVT and NVE ensembles. A key contribution of the paper is the PME algorithm that we redesigned from a charge-centric to a lattice-of-points-centric prospective to better fit the multithreading GPU architecture. FENZI enables multi-nanosecond MD simulations per day, even when using PME for large membranes. FENZI achieves simulation rates of up to 22.8 nanoseconds per day with a MD step size of 1 fs for a small DMPC membrane of 17,004 atoms, up to 6.7 nanoseconds per day with a MD step size of 1 fs for a medium DMPC membrane of 68,484 atoms, and up to 1.6 nanoseconds per day with a MD step size of 1 fs for a large DMPC membrane of 273,936 atoms; all membranes were in explicit solvent. Because of our general design and implementation approach, FENZI is full-fledged general purpose MD package based on canonical force fields and PME for GPU enabled simulation of a broad class of molecular systems.
Work in progress includes the analysis of properties such as structural (densities, electron density profiles), order parameters (SCD) and electrostatic properties (dipole potential, water dipole moments), as well as orientational properties of water over a time scale of tens of hundreds nanoseconds. Our final goal is to assess whether properties observed in small membranes with simulations on traditional CPU clusters are still true as the membranes significantly grow in size. By speeding up the MD simulation of large membranes, FENZI allows us to answer this question accurately in a short turnaround time.
Ganesan, N.; Taufer, M.; Bauer, B.; Patel, S.; , “FENZI: GPU-Enabled Molecular Dynamics Simulations of Large Membrane Regions Based on the CHARMM Force Field and PME,” Parallel and Distributed Processing Workshops and Phd Forum (IPDPSW), 2011 IEEE International Symposium on , vol., no., pp.472-480, 16-20 May 2011 [doi: 10.1109/IPDPS.2011.187] [Free PDF]