使用GPU加速分子动力学模拟中的非绑定力计算

在分子动力学模拟(MD)中,对非绑定力的计算需要花费大量的时间。本文提出了基于CUDA和Brook+的两种双精度算法,分别在NVIDIA和AMD两款主流GPU上实现了非绑定力的计算,借助GPU的计算能力加速了整个MD程序。算法对MD进行了任务分割,采用区域分解的方法将非绑定力的计算映射到GPU的计算核心上,同时针对两款GPU的各自特点提出了线程块内共享存储、最小化数据集两种优化方法。性能测试结果表明,与Intel Xeon 2.6GHzCPU的单核相比,43.2万粒子的高速粒子碰撞模拟,在配置NVIDIA Tesla C1060的系统上性能提高了6.5倍,在配置AMD HD4870的系统上性能提高了4.8倍。

GPU Acceleration of Nonbonded Forces for Molecular Dynamics Simulation

The most costly computation required by molecular dynamics (MD) is the calculation of non-bonded forces. This paper presents two algorithms based on CUDA and Brook+ to accelerate the calculation of non-bonded forces and then examines them on two popular GPUs of NVIDIA and AMD. The algorithms enable fine-grained task decomposition and spatial decomposition of forces that map efficiently to the compute units of GPU. They also present optimization methods such as use of shared memory between threads and strip-mined. We then test our implementations for 432 thousand atoms. Our code executed on Tesla C1060 runs 6.5 times faster than the original application running on single core of Intel Intel Xeon 2.6GHZ CPU, while on HD4870 it runs 4.8 times faster.