摘要
ABEEMσπ(Atom-Bond Electronegativity EqualizationσπModel)模型中,原串行程序求静电相互作用能的方法非常耗时,致使研究问题的效率降低。针对原程序中多个循环相互嵌套的求解部分,采用带状卷帘存储迭代分配的MPI(Message Passing Inter-face)并行化处理;对体系中所有原子、σ键、孤对电子、π键位点之间的静电相互作用能采用多线程CUDA(Computer Unified DeviceArchitecture)并行化处理。传统MPI+CUDA环境中,GPU和CPU之间的数据传输开销大,导致整体性能下降以及各种粒子间计算串行调用CUDA,致使时间浪费。针对上述情况,使用GPU核心的缓存机制解决传输开销大的问题,并利用多CUDA流技术实现多个循环异步进行计算,从而缩短了运行时间。然后选取多个不同类型的大分子体系进行测试,结果表明,利用改进的MPI+CUDA并行模型进行动力学模拟,并行加速比显著提高,大幅度缩减了求解静电相互作用能的时间,并得到与串行一致的结果。
In ABEEMσπ model, original serial program consumes much time in seeking electrostatic interaction energy, which caused the research inefficient. In solution part of the original program, as the multiple loops are nested each other, MPI parallel processing of strip rolling storage iterative distribution is adopted to resolve this problem; and, multi-threaded CUDA parallel processing is used to deal with the static sites interactions among all the atoms, σ bond, lone pair electrons and ,π bond in the system. In traditional environment of MPI ＋ CUDA, there is huge spending when data transferring between GPU and CPU, which results in overall performance decrease and the calculation of serial called CUDA between a variety of particles, therefore leads to time wasted. For these above, this paper proposes that applying the mechanism of the GPU core caching to solve the problem of huge transmission cost, and making use of multi-stream technology of CUDA to realise multiple cycles asynchronous for calculation, so that the running time will be reduced. Then, several systems of different types of macromolecular are selected to test, the result shows, by applying modified MPI ＋ CUDA parallel model in dynamics simulation, the parallel speedup improves significantly, the time of solving the electrostatic interaction energy reduces substantially, while results are identical to the serial program.
出处
《计算机应用与软件》
CSCD
北大核心
2012年第11期35-38,共4页
Computer Applications and Software
基金
国家自然科学基金项目(21133005
20703022
21011120087)
