G8 Multilateral Research Funding - ICOMEX

目前仍需要具备以更高分辨率运行大气模式的能力，以改进天气预报，特别是恶劣天气预报，提高对气候过程的了解，并改进气候预测。更高的分辨率只能通过利用基于数十万处理器的未来超级计算平台来实现；对于这样的架构，处理器之间的通信是性能的重要瓶颈，并且需要与历史上使用的算法完全不同的算法。基于精细二十面体的网格为新一代大气模式提供了有希望的基础。它们的准均匀分辨率可以最大限度地减少通信瓶颈，从而提供良好的性能。数值分析的最新进展意味着现在有足够精确的方案。在特定应用需要更高分辨率的地方，可以对网格进行局部细化。一个二十面体网格模型已经被德国气象局应用了好几年，其他几个小组也有这样的模型处于发展的高级阶段。该项目将评估和比较多个二十面体网格模型的并行计算性能，并开展改进其性能的研究。具体的研究课题包括开发一种抽象模型描述方案，以尽可能地将数值分析问题与计算机科学问题分开，针对基于图形处理单元（GPU）的架构的模型实现，开发和评估椭圆求解器以确定稳定的长时间步长积分方案的可行性，开发内部后处理包，以提供用户所需的诊断，同时最大限度地减少数据输出瓶颈，并制定优化输入/输出的策略。埃克塞特大学将特别致力于椭圆求解器和后处理工作包。

There is an ongoing need for a capability to run atmospheric models at higher resolution, to improve weather forecasts, particularly of severe weather, to improve understanding of climate processes, and to improve climate prediction. Higher resolution can only be achieved by exploiting future supercomputing platforms based on hundreds of thousand of processors; for such architectures communication between processors is a significant bottleneck to performance, and requires quite different algorithms from those used historically.Grids based on a refined icosahedron provide a promising basis for a new generation of atmospheric models. Their quasi-uniform resolution should minimize the communication bottleneck giving good performance. Recent advances in numerical analysis mean that schemes of sufficient accuracy are now available. Local refinement of the grid is possible where higher resolution is needed for specific applications. An icosahedral grid model has been used operationally by the German Weather Service for several years, and several other groups have such models at an advanced stage of development.The proposed project will assess and compare the parallel computing performance of a number of icosahedral grid models and carry out research to improve their performance. The specific research topics include the development of an abstract model description scheme to separate (as far as possible) numerical analysis issues from computer science issues, a model implementation targeted on Graphical Processing Unit (GPU) based architectures, development and evaluation of elliptic solvers to determine the feasibility of stable long-time-step integration schemes, the development of an internal postprocessing package to provide diagnostics required by the user while minimizing the data output bottleneck, and development of strategies to optimize input/output. The University of Exeter will contribute particularly to the elliptic solver and postprocessing work packages.