ACEMD: Accelerating bio-molecular dynamics in the microsecond time-scale

📝 Abstract

The high arithmetic performance and intrinsic parallelism of recent graphical processing units (GPUs) can offer a technological edge for molecular dynamics simulations. ACEMD is a production-class bio-molecular dynamics (MD) simulation program designed specifically for GPUs which is able to achieve supercomputing scale performance of 40 nanoseconds/day for all-atom protein systems with over 23,000 atoms. We illustrate the characteristics of the code, its validation and performance. We also run a microsecond-long trajectory for an all-atom molecular system in explicit TIP3P water on a single workstation computer equipped with just 3 GPUs. This performance on cost effective hardware allows ACEMD to reach microsecond timescales routinely with important implications in terms of scientific applications.

💡 Analysis

The high arithmetic performance and intrinsic parallelism of recent graphical processing units (GPUs) can offer a technological edge for molecular dynamics simulations. ACEMD is a production-class bio-molecular dynamics (MD) simulation program designed specifically for GPUs which is able to achieve supercomputing scale performance of 40 nanoseconds/day for all-atom protein systems with over 23,000 atoms. We illustrate the characteristics of the code, its validation and performance. We also run a microsecond-long trajectory for an all-atom molecular system in explicit TIP3P water on a single workstation computer equipped with just 3 GPUs. This performance on cost effective hardware allows ACEMD to reach microsecond timescales routinely with important implications in terms of scientific applications.

📄 Content

arXiv:0902.0827v1 [physics.comp-ph] 5 Feb 2009 ACEMD : Accelerating bio-molecular dynamics in the microsecond time-scale M. J. Harvey,1, ∗G. Giupponi,2, † and G. De Fabritiis3, ‡ 1Information and Communications Technologies, Imperial College London, South Kensington, London, SW7 2AZ, UK 2Department de Fisica Fundamental, Universitat de Barcelona, Carrer Marti i Franques 1, 08028 Barcelona, Spain 3Computational Biochemistry and Biophysics Lab (GRIB-IMIM), Universitat Pompeu Fabra, Barcelona Biomedical Research Park (PRBB), C/ Doctor Aiguader 88, 08003 Barcelona, Spain Abstract The high arithmetic performance and intrinsic parallelism of recent graphical processing units (GPUs) can offer a technological edge for molecular dynamics simulations. ACEMD is a production-class bio-molecular dynamics (MD) simulation program designed specifically for GPUs which is able to achieve supercomputing scale performance of 40 nanoseconds/day for all-atom pro- tein systems with over 23, 000 atoms. We illustrate the characteristics of the code, its validation and performance. We also run a microsecond-long trajectory for an all-atom molecular system in explicit TIP3P water on a single workstation computer equipped with just 3 GPUs. This perfor- mance on cost effective hardware allows ACEMD to reach microsecond timescales routinely with important implications in terms of scientific applications. 1 I. INTRODUCTION The simulation of mesoscopic scales (microseconds to milliseconds) of macromolecules continues to pose a challenge to modern computational biophysics. Whilst the fundamental thermodynamic framework behind the simulation of macromolecules is well characterised, exploration of biological time scales remains beyond the computational capacity routinely available to many researchers. This has significantly inhibited the widespread use of molec- ular simulations for in silico modelling and prediction1. Recently, there has been a renewed interest in the development of molecular dynamics simulation techniques. D. E. Shaw Research2 has fostered several significant algorithmic improvements including mid-point3 and neutral-territory methods4 for the summation of non-bonded force calculations, a new molecular dynamics package called Desmond5 and Anton2, a parallel machine for molecular dynamics simulations that uses specially-designed hardware. Other parallel MD codes, such as Blue matter6, NAMD7 and Gromacs48, have been designed to perform parallel MD simulations across multiple independent processors, but latency and bandwidth limitations in the interconnection network between processors reduces parallel scaling unless the size of the simulated system is increased with proces- sor count. Furthermore, dedicated, highly-parallel machines are usually expensive and not reservable for long periods of time due to cost constraints and allocation restrictions. A further line of development of MD codes consists of using commodity high perfor- mance accelerated processors1. This approach has become an active area of investigation, particularly in relation to the Sony-Toshiba-IBM Cell processor9 and graphical processing units (GPUs). Recently, De Fabritiis9 implemented an all-atom biomolecular simulation code, CellMD, targeted to the architecture of the Cell processor (contained within the Sony Playstation3) that reached a sustained performance of 30 Gflops with a speedup of 19 times compared to the single CPU version of the code. At the same time, a port of the Gromacs code for implicit solvent models10 was developed and used by the Folding@home distributed computing project11 on a distributed network of Playstation3s. Similarly, CellMD was used in the PS3GRID.net project12 based on the BOINC platform13 moving all-atom MD appli- cations into a distributed computing infrastructure. Pioneers in the use of GPUs for production molecular dynamics11 had several limita- tions imposed by the restrictive, graphics-orientated OpenGL programming model14 then 2 available. In recent years, commodity GPUs have acquired non-graphical, general-purpose programmability and have undergone a doubling of computational power every 12 months, compared to 18 −24 months for traditional CPUs1. Of the devices currently available on the market, those produced by Nvidia offer the most mature programming environment, the so-called compute unified device architecture (CUDA)15, and have been the focus of the majority of investigation in the computational science field. Several groups have lately shown results for MD codes which utilise CUDA-capable GPUs. Stone et al 16 demonstrated the GPU-accelerated computation of the electrostatic and van der Waals forces, reporting a 5.4 times speed-up with respect to a conventional CPU. Meel et al 17 described an implementation for simpler Lennard-Jones atoms which achieved a net speedup of up to 40 times over a conventional CPU. Unlike the former, the whole simulation is performed on the GPU. Recently, Phillips et al have reported experim

This content is AI-processed based on ArXiv data.