摘要
分子动力学模拟(Molecular Dynamics,MD)是计算化学和生物模拟领域一种重要的计算手段,由于计算强度大,目前MD可模拟的时空尺度还不能满足真实物理过程的需要,计算速度是其主要瓶颈之一。2007年以来,比CPU具有更强大的存储器带宽和计算能力的GPU(Graphics Processing Units)的可编程能力获得了显著提升,为数值计算的并行加速提供了一种新的选择。除了使用并行技术加速MD,合理地使用约束算法可增大模拟的时间步长以降低MD计算量。本文首次建立了GPU加速的LINCS(Linear Constraint Solver)约束算法GMD_LINCS,使用线程组织、合并访问、全局同步等对其进行了优化。GMD_LINCS是基于GPU的MD程序(GMD)的约束算法部分。采用GROMACS官网提供的基准算例二氢叶酸还原酶(DHFR)对GMD_LINCS的测试结果表明,GMD_LINCS程序和GROMACS4.5.3CPU版本的计算精度吻合较好。对含有19万个粒子(27条链)的聚丙烯腈(PAN)算例的测试结果表明,GMD_LINCS程序的计算性能获得明显提升,比GROMACS4.5.3相应的LINCS约束算法的单核CPU性能可加速约17倍、是其八核CPU性能的4.5倍左右。
Molecular dynamics（MD） is an important computing method in computational chemistry and biomolecules simulation. Because of its high computation costs, MD simulations currently are still limited in temporal and spatial scales to meet the real world applications. In recent years, the rapid improvements of graphic processing units（GPU） for general purpose programming with its powerful memory bandwidth and computing ability over CPU offer an alternative for numerical calculation in parallel. In addition to parallel computing, the performance of MD simulation can be improved using constraint algorithms appropriately to reduce the amount of MD calculation by increasing the time step. In this paper, we present GMD-L1NCS, the first implementation of GPU-enabled linear constraint solver（LINCS）, and its optimization strategies in thread organization, memory coalescence and global synchronization. GMD LINCS is the constraint algorithms adopted in GMD, a GPU-based MD program. GMD L1NCS has been benchmarked via dihydrofolate reductase （DHFR） from GROMACS official website and results show that it has good agreement with GROMACS 4.5.3 CPU version in calculation precision. In addition, the test data for 190,000 particles system comprising 27 identical polyacrylonitrile （PAN） chains show that GMD LINCS obtains good performance with speedup of 17 times over the LINCS in GROMACS 4.5.3 for single core CPU and 4.5 times over its eight core CPU performance respectively.
出处
《计算机与应用化学》
CAS
CSCD
北大核心
2012年第8期907-912,共6页
Computers and Applied Chemistry
基金
国家自然科学基金资助项目(21073195)
中国科学院过程工程研究所多相复杂系统国家重点实验室开放基金资助
作者简介
刘忠亮（1987-），男，山东人，硕士研究生，Email：zhongliangv@gmail．com.
联系人：李晓霞（1964-），女，山西人，研究员，Email：xxia@home．ipe．ac．cn
