ＧＰＵによって加速化された氷床モデリング用プログラムの開発と応用

计算机模拟已成为了解冰盖动力学的一项重要技术。特别是，它们与预测未来气候变化中冰盖的演变有关。该项目的目标是开发一个可操作的冰盖模型，该模型采用GPU（图形处理单元）加速计算技术。这个模型被称为“Sainou”，是一个Elmer/Ice （elmerice.elmerfem.org）的导数，它用有限元方法解决了冰动力学的全斯托克斯方程（不忽略应力）。在第二个项目年度（2016财年），重点是对Sainou进行测试，使用已建立的基准问题，即ISMIP-HOM实验A （Pattyn et al., 2008, Cryosphere 2,95 -108）。这个实验包括一个倾斜的冰板流过一个凹凸不平的床。将Sainou对全- stokes偏微分方程系统的有限元矩阵的装配所需的计算时间与Elmer/Ice所需的计算时间进行了比较。在ISMIP-HOM实验A中测试了四种不同数量的元素，并在Mac Pro工作站上进行了模拟，仅使用单个CPU。结果发现，对于双精度精度，分别为8000、87500、252000和391500块砖，Sainou比Elmer/Ice快7倍、17倍、23倍和25倍，对于单精度精度，Sainou比Elmer/Ice快9倍、38倍、45倍和49倍。因此，可以成功地证明gpu加速计算是一种快速有效的冰盖建模方法。

Computer simulations have become established as an important technique for the understanding of ice sheet dynamics. In particular, they are relevant for predicting the evolution of ice sheets in a future changing climate. The goal of this project was to develop an operational ice sheet model that employs the technique of GPU- (graphics processing unit) accelerated computing.This model, termed "Sainou", is an Elmer/Ice (elmerice.elmerfem.org) derivative that solves the full-Stokes equations (no stresses neglected) of ice dynamics with the finite element method. In the second project year (FY 2016), the focus was on testing Sainou, using an established benchmark problem, namely the ISMIP-HOM experiment A (Pattyn et al., 2008, Cryosphere 2, 95-108). This experiment consists of an inclined slab of ice flowing over a bumpy bed.The computing time required by Sainou for the assembly of the finite element matrix of the full-Stokes system of partial differential equations was compared to that required by Elmer/Ice. Four different numbers of elements were tested for the ISMIP-HOM experiment A, and the simulations were carried out on a Mac Pro workstation, using only a single CPU. It was found that, for double-precision accuracy with 8000, 87500, 252000 and 391500 brick elements, respectively, Sainou is 7x, 17x, 23x and 25x faster than Elmer/Ice, and for single-precision accuracy it is even 9x, 38x, 45x and 49x faster. It could therefore be demonstrated successfully that GPU-accelerated computing is a useful method for fast and efficient ice sheet modelling.