针对具有物理机制的分布式水文模型对大流域、长序列模拟计算时间长、模拟速度慢的问题,引入基于GPU的并行计算技术,实现分布式水文模型WEP-L(water and energy transfer processes in large river basins)产流过程的并行化。选择鄱阳...针对具有物理机制的分布式水文模型对大流域、长序列模拟计算时间长、模拟速度慢的问题,引入基于GPU的并行计算技术,实现分布式水文模型WEP-L(water and energy transfer processes in large river basins)产流过程的并行化。选择鄱阳湖流域为实验区,采用计算能力为8.6的NVIDIA RTX A4000对算法性能进行测试。研究表明:提出的基于GPU的分布式水文模型并行算法具有良好的加速效果,当线程总数越接近划分的子流域个数(计算任务量)时,并行性能越好,在实验流域WEP-L模型子流域单元为8712个时,加速比最大达到2.5左右;随着计算任务量的增加,加速比逐渐增大,当实验流域WEP-L模型子流域单元增加到24897个时,加速比能达到3.5,表明GPU并行算法在大尺度流域分布式水文模型计算中具有良好的发展潜力。展开更多
Designing and optimizing complex scientific code for new computing architectures is a challenging task. To address this issue in the E3SM land model (ELM) development, we developed a software tool called SPEL, which f...Designing and optimizing complex scientific code for new computing architectures is a challenging task. To address this issue in the E3SM land model (ELM) development, we developed a software tool called SPEL, which facilitates code generation, verification, and performance tuning using compiler directives within a Function Unit Test framework. In this paper, we present a SPEL extension that leverages the version control system (e.g., Git) to autonomous code generation and demonstrate its application to continuous code integration and development of the ELM software system. The study can benefit the scientific software development community.展开更多
针对深度学习图像分类场景中多GPU并行后传输效率低的问题,提出一种低时间复杂度的Ring All Reduce改进算法。通过分节点间隔配对原则优化数据传输流程,缓解传统参数服务器并行结构的带宽损耗。基于数据并行难以支撑大规模网络参数及加...针对深度学习图像分类场景中多GPU并行后传输效率低的问题,提出一种低时间复杂度的Ring All Reduce改进算法。通过分节点间隔配对原则优化数据传输流程,缓解传统参数服务器并行结构的带宽损耗。基于数据并行难以支撑大规模网络参数及加速延缓的问题,根据深度学习主干网络所包含的权重参数低于全连接层权重参数、同步开销小、全连接层权重大与梯度传输开销过高等特点,提出GPU混合并行优化算法,将主干网络进行数据并行,全连接层进行模型并行,并通过改进的Ring All Reduce算法实现各节点之间的并行后数据通信,用于基于深度学习模型的图像分类。在Cifar10和mini ImageNet两个公共数据集上的实验结果表明,该算法在保持分类精度不变的情况下可以获得更好的加速效果,相比数据并行方法,可达到近45%的提升效果。展开更多
文摘针对具有物理机制的分布式水文模型对大流域、长序列模拟计算时间长、模拟速度慢的问题,引入基于GPU的并行计算技术,实现分布式水文模型WEP-L(water and energy transfer processes in large river basins)产流过程的并行化。选择鄱阳湖流域为实验区,采用计算能力为8.6的NVIDIA RTX A4000对算法性能进行测试。研究表明:提出的基于GPU的分布式水文模型并行算法具有良好的加速效果,当线程总数越接近划分的子流域个数(计算任务量)时,并行性能越好,在实验流域WEP-L模型子流域单元为8712个时,加速比最大达到2.5左右;随着计算任务量的增加,加速比逐渐增大,当实验流域WEP-L模型子流域单元增加到24897个时,加速比能达到3.5,表明GPU并行算法在大尺度流域分布式水文模型计算中具有良好的发展潜力。
文摘Designing and optimizing complex scientific code for new computing architectures is a challenging task. To address this issue in the E3SM land model (ELM) development, we developed a software tool called SPEL, which facilitates code generation, verification, and performance tuning using compiler directives within a Function Unit Test framework. In this paper, we present a SPEL extension that leverages the version control system (e.g., Git) to autonomous code generation and demonstrate its application to continuous code integration and development of the ELM software system. The study can benefit the scientific software development community.
文摘针对深度学习图像分类场景中多GPU并行后传输效率低的问题,提出一种低时间复杂度的Ring All Reduce改进算法。通过分节点间隔配对原则优化数据传输流程,缓解传统参数服务器并行结构的带宽损耗。基于数据并行难以支撑大规模网络参数及加速延缓的问题,根据深度学习主干网络所包含的权重参数低于全连接层权重参数、同步开销小、全连接层权重大与梯度传输开销过高等特点,提出GPU混合并行优化算法,将主干网络进行数据并行,全连接层进行模型并行,并通过改进的Ring All Reduce算法实现各节点之间的并行后数据通信,用于基于深度学习模型的图像分类。在Cifar10和mini ImageNet两个公共数据集上的实验结果表明,该算法在保持分类精度不变的情况下可以获得更好的加速效果,相比数据并行方法,可达到近45%的提升效果。
基金Project(52078498)supported by the National Natural Science Foundation of ChinaProject(2021-Special-04-2)supported by the Science and Technology Research and Development Program Project of China Railway Group Limited+2 种基金Project(2022JJ30745)supported by the Natural Science Foundation of Hunan Province,ChinaProject(2023QYJC006)supported by the Frontier Cross Research Project of Central South University,ChinaProject(2020TJQ19)supported by the Hunan Provincial Science and Technology Promotion Talent Project,China。
