Acceleration of Biomolecular Simulations using FPGA-based Reconfigurable Computing

A paradigm shift is occurring in the way compute-intensive scientific applications are developed. Thanks to advancements in commercially viable hybrid architectures for High-Performance Computing (HPC), the focus has shifted from improving performance by merely scaling algorithms on von Neumann computing nodes to fully exploiting additional computational capabilities provided by accelerators such as FPGAs (Field Programmable Gate Arrays) and GPGPUs (General Purpose Graphical Processing Units). Computational chemists use Molecular Dynamics (MD) simulations like LAMMPS (Large Scale Atomic Molecular Massively Parallel Systems) and NAMD (NAnoscale Molecular Dynamics) to simulate biomolecular behaviour such as protein folding and small molecule docking to proteins. MD simulations are computationally complex n-body problems, which are time consuming to simulate in biologically relevant scales. Executing such simulations in best available HPC environments is critical for scientific advancements in the field. Thus, as HPC technology evolves, there is a need to update classical biomolecular simulation applications like LAMMPS to better suit the architecture. In this work, we modify LAMMPS (a classical molecular dynamics simulation program developed for CPU-only clusters) to execute on a reconfigurable computer system, SRC-7 H MAP. The SRC-7 H MAP consists of two Altera FPGA logic chips interfaced to a dual-core Intel Xeon processor. Users can benefit by offloading most compute-intensive tasks of the application to the FPGA logic. This work explores the challenges involved in effectively adapting a production level application code optimized for von Neumann architecture, to an FPGA-based hybrid architecture. We have successfully accelerated the non-bonded force computations, the most compute-intensive module in LAMMPS for biomolecular simulations, by 5.0x over a single 3.0 GHz Xeon processor. This performance includes the data transfer overheads and function calling overheads. Further, using the accelerated non-bonded force computations function, we achieve an overall application speed-up of 2.0x to 2.4