Accelerating molecular dynamics simulations with high performance reconfigurable systems.
详细信息
文摘
Molecular Dynamics MD) simulation plays an important role in understanding functionally important phenomena of biological processes. Many of such events,however,occur on time scales beyond the capabilities of modern computers. Hence,accelerating MD simulations becomes one of critical challenges in the current study of Computational Biology and Chemistry. For the past several years,various implementations and efficient algorithms have been developed toward this goal. Of these,accelerating MD with Field Programmable Gate Arrays FPGAs) has been shown to be a viable candidate for improved MD cost/performance. Given the intense competition from multi-core and GPUs,there is now a question whether MD on High Performance Reconfigurable Computing HPRC) can be competitive. The goal of this research is to create an FPGA-based MD system to achieve substantial speedup over production MD software without compromising simulation quality. In one part of the study,we systematically explore and evaluate the design space of the force pipeline with respect to arithmetic algorithm,arithmetic mode,precision,equation complexity,and various other optimizations. We find that the FPGAs Block RAM BRAM) architecture makes them well suited to support unusually fine-grained intervals. This leads to a reduction in other logic and a proportional increase in performance. In the other part,we present the first FPGA study of the filtering of particle pairs having nearly zero mutual force,a standard optimization in MD codes. There are several innovations,including a novel partitioning of the particle space,and new methods for filtering and mapping work onto the pipelines. As a consequence,highly efficient filtering can be implemented with only a small fraction of the FPGAs resources. Overall,we find that,for an Altera Stratix-III EP3ES260,six force pipelines running at nearly 200 MHz can fit on the FPGA. This results in a 26x per core speedup for the nonbonded short-range force. The resulting integrated system is likely to make FPGAs highly competitive for MD.