【作者】 李鸿健；

【导师】 孙世新；

【作者基本信息】 电子科技大学 ， 计算机软件与理论， 2012， 博士

【摘要】 随着超短激光脉冲技术的不断发展，激光化学反应的研究已经深入到核运动的微观世界。许多化学反应实验采用常规方法无法实现，但在超快激光脉冲的照射下能够实现。激光化学反应模拟是利用计算机手段模拟分子在不同激光照射下发生的化学反应现象，计算机模拟能够得到激光化学反应全过程，研究瞬间反应的具体细节，为实际化学反应提供理论指导。目前应用最为广泛的激光化学反应模拟方法是半经典分子动力学方法，由于当模拟规模增加时，其计算量迅速增大，因此，研究高效率并行算法并应用于半经典动力学模拟具有重要意义。本文全面系统地研究激光化学反应模拟并行算法，在此基础上设计了一系列并行算法。本文的主要内容及创新点如下：1．本文基于半经典分子动力学模型实现激光化学反应模拟。该模型由Fortran语言实现，并在大型计算机上构建模拟平台。该平台基于半经典的计算机模拟激光化学反应模型，能够真实反映化学反应的实际过程。通过实际的激光化学反应的模拟测试，实验结果证明该平台能够有效研究激光脉冲性质对化学反应的影响，为研究激光化学反应提供了科学可靠方法。2．本文提出了半经典分子动力学特殊矩阵计算的优化方法和并行计算方法。通过对半经典分子动力学模拟计算的串行算法进行了测试分析，寻求耗时最大的程序模块，即大型特殊矩阵乘法计算。基于计算矩阵特征对特殊矩阵乘法进行优化，大幅降低该模块计算时间。采用Wingrad算法降低矩阵计算时间复杂度，并基于OpenMP实现矩阵乘的Winograd并行算法，有效提高计算效率。3．本文提出了一种半经典分子动力学混合并行算法。在半经典分子动力学模拟中引入混合并行技术和双层并行算法设计。基于MPI+OpenMP混合模型和激光化学反应计算特征，设计并实现激光化学反应模拟双层并行算法，上层并行采用动态原子分割算法，下层并行采用力矩阵并行分解算法。该算法在多核集群中能够有效提高半经典分子动力学并行计算效率。4．基于反馈机制设计动态负载均衡算法，该算法通过负载信息反馈动态调整任务分配，并将该算法应用与半经典分子动力学模拟中，在变化负载系统中，该算法具有较好的适应性。为了进一步适应复杂的频繁变化的负载系统，设计了可变周期反馈的负载均衡算法，根据实时负载调整反馈周期，从而增强算法的可适应性，负载变化较多时能缩短反馈周期，负载稳定时能延长周期，从而提高系统并行效率。针对半经典分子动力学模拟计算特征，设计了周期时间预测算法和优化原子分配方案，在多核集群中测试，采用随机运行的干扰程序动态改变系统负载，在负载动态变化环境中，该负载均衡算法具有优越的自适应性和可扩展性，能够有效提高系统整体性能。更多还原

【Abstract】 The development of ultrashort laser pulse allows us to study photochemicalreactions at atomic level. Many chemical reactions that can not be realized bytraditional techniques become possible due to ultrashort laser technology. The mostpopular technique for studying laser induced chemical reactions is Pump-Probespectroscopy. Computer simulation of photochemical reactions is a complementaryapproch to the investigations of the photochemical reactions. Following aphotochemical reaction using computer simulation, researchers can view the chemicalreaction from reactants to products and gain the information about reactionmechanisms.1.A model for realistic parallel simulations of photochemical reactions induced byan ultrashort laser pulses is developed and realized by Fortran90language. Theinterreaction between laser pulses and molecules is unambiguously included in themodel. The technique allows researchers to investigate the influence of the properties oflaser pulses on the outcome of chemical reactions and therefore is a realistic approach tothe study of laser control of chemical reactions. The efficiency of the program isimproved by parallel design, It provides a practical technique for performingphotochemical reactions starting from laser excitation high efficiently. In addition, itprovides many details that could not be observed experimentally.2.An optical algorithm of special matrix multiplication is suggested in semiclassicalmolecular dynamics simulations. Based on the analysis of the serial program, the matrixmultiplication cost most time of the force calculation in semiclassical moleculardynamics simulations. The computation of these gradient matrices and theirmultiplication plays a key role in the Ehrenfest force calculation. Simplifying and thenparallelizing the computation of these nuclear gradients and the matrix multiplicationsnumerically eventually enhance the capacity of Ehrenfest MD simulation.3.A high performance hybrid parallel algorithm for simulating photochemicalreaction is developed by introducing the concept of two layers of parallel capacity. MPIis used in conjunction with OpenMP multithreading to achieve two level parallism of the data and tasks. It is realized based on SMP cluster. It is proved that this method is afeasible parallel algorithm for simulating the photochemical reactions high-efficiently.4.To improve the parallel efficiency on unbalancing system, A dynamic loadbalancing algorithm is designed based on the feedback model. All Tasks areredistributed by load feedback and processing ability forecast. This algorithm adoptedthe dynamic variation of the load. In this study, the algorithm is applied andimplemented in the force calculation of semiclassical molecular dynamic simulations,an optical atomic decomposition is also implemented. Compare with the static loadbalance algorithm, this dynamic load balance algorithm shows better adaptability andscalability.更多还原