Molecular dynamics simulations offer the possibility to understand a wide range of biological activity and estimate equilibrium and dynamic properties of complex systems that cannot be calculated analytically or observed empirically. By simulating the motion of a molecular system in space and time, we can observe its wide range of thermally-accessible states, and eventually understand the connection between the system’s structure and its function. Indeed, molecular dynamics simulations have already proved successful in commercial examples of complex drug design.
In this line of research we collaborate with Plebiotic
PleMD Molecular Dynamics Platform.
PleMD is a molecular dynamics platform developed jointly with and commercialized by Plebiotic. PleMD implements a highly efficient algorithm for molecular dynamics simulation that exploits the hybrid multi-CPU multi-GPU computing capabilities of today’s desktop computers. It relies on top-notch numerical methods and data structures to maximize the simulation time step while minimizing the operation count. But, most importantly, these algorithms are parallelized and distributed over a hybrid multi-CPU multi-GPU architecture to balance the computational load over the available resources and maximize performance.
Multi-GPU Molecular Dynamics.
Molecular dynamics simulations allow us to study the behavior of complex biomolecular systems. These simulations suffer a large computational complexity that leads to simulation times of several weeks in order to recreate just a few microseconds of a molecule’s motion even on high-performance computing platforms. In recent years, state-of-the-art molecular dynamics algorithms have benefited from the parallel computing capabilities of multicore systems, as well as GPUs used as co-processors.
We are working on a parallel molecular dynamics algorithm for on-board multi-GPU architectures. We parallelize a state-of-the-art molecular dynamics algorithm at two levels. We employ a spatial partitioning approach to simulate the dynamics of one portion of a molecular system on each GPU, and we take advantage of direct communication between GPUs to transfer data among portions. We also parallelize the simulation algorithm to exploit the multi-processor computing model of GPUs.
Following these guidelines, we developed a parallel and scalable solution to compute bond-forces, short-range forces and long-range molecular forces (based on the multilevel summation method (MSM)). We developed a novel parallel algorithms to update the spatial partitioning and set up transfer data packages on each GPU. We demonstrate the feasibility and scalability of our proposal through a comparative study with NAMD, a well known parallel molecular dynamics implementation.
Our implementation for long-range molecular forces computation demonstrate an optimization of MSM that replaces 3D convolutions with FFTs, and we achieve a single-GPU performance comparable to the particle mesh Ewald (PME) method, the de facto standard for long-range molecular force computation. But most importantly, we implemented a distributed version of MSM that avoids the scalability difficulties of PME.