在高性能计算领域,拥有强大浮点计算能力的协处理器正在快速发展。近年来,利用协处理器(如GPU)来加速时域有限差分FDTD算法的计算过程成为电磁研究领域的热点问题。在GPU集群上实现了三维UPML-FDTD算法并进行了优化。采用电偶极子激励...在高性能计算领域,拥有强大浮点计算能力的协处理器正在快速发展。近年来,利用协处理器(如GPU)来加速时域有限差分FDTD算法的计算过程成为电磁研究领域的热点问题。在GPU集群上实现了三维UPML-FDTD算法并进行了优化。采用电偶极子激励源对算法的模拟结果同解析解进行了验证,结果表明该算法具有较高的精度;同时,在NVIDIA Tesla M2070和K20mGPU集群上对FDTD算法的性能进行测试,对优化前后的计算结果以及GPU与CPU的计算性能进行了比较,并使用80块NVIDIA Tesla K20mGPU进行了可扩展性测试。从本文的研究结果可以看出,经过优化的FDTD算法性能有了较大的提升,而且FDTD算法在GPU集群上获得了比较理想的并行效率。展开更多
We utilize Fourier methods to analyze the stability of the Yee difference schemes for Berenger PML (perfectly matched layer) as well as the UPML (uniaxial perfectly matched layer) systems of two-dimensional Maxwel...We utilize Fourier methods to analyze the stability of the Yee difference schemes for Berenger PML (perfectly matched layer) as well as the UPML (uniaxial perfectly matched layer) systems of two-dimensional Maxwell equations. Using a practical spectrum stability concept, we find that the two schemes are spectrum stable under the same conditions for mesh sizes. Besides, we prove that the UPML schemes with the same damping in both directions are stable. Numerical examples are given to confirm the stability analysis for the PML method.展开更多
文摘在高性能计算领域,拥有强大浮点计算能力的协处理器正在快速发展。近年来,利用协处理器(如GPU)来加速时域有限差分FDTD算法的计算过程成为电磁研究领域的热点问题。在GPU集群上实现了三维UPML-FDTD算法并进行了优化。采用电偶极子激励源对算法的模拟结果同解析解进行了验证,结果表明该算法具有较高的精度;同时,在NVIDIA Tesla M2070和K20mGPU集群上对FDTD算法的性能进行测试,对优化前后的计算结果以及GPU与CPU的计算性能进行了比较,并使用80块NVIDIA Tesla K20mGPU进行了可扩展性测试。从本文的研究结果可以看出,经过优化的FDTD算法性能有了较大的提升,而且FDTD算法在GPU集群上获得了比较理想的并行效率。
文摘We utilize Fourier methods to analyze the stability of the Yee difference schemes for Berenger PML (perfectly matched layer) as well as the UPML (uniaxial perfectly matched layer) systems of two-dimensional Maxwell equations. Using a practical spectrum stability concept, we find that the two schemes are spectrum stable under the same conditions for mesh sizes. Besides, we prove that the UPML schemes with the same damping in both directions are stable. Numerical examples are given to confirm the stability analysis for the PML method.