A scalable dynamical core for Next Generation Weather and Climate Prediction - Phase 2

从历史上看，数值天气预报和气候模拟精度的主要改进来自于计算机能力指数级增长所带来的分辨率的提高。为了通过进一步提高分辨率来实现精度的进一步提高，有必要利用正在变得可用的大规模并行计算机架构。然而，目前最先进的运算算法并不期望在几千个处理器之外表现良好：传统经纬度网格的网格结构意味着处理器间通信最终但不可避免地成为瓶颈。拟议项目的总体目标是开发一种新的、三维的、完全可压缩的动力核心，适用于大规模并行机器上的全球和区域天气和气候预测，以及研究用途，并证明其准确性、效率和可扩展性。准确度应与现有的最先进的算法相媲美。该算法必须足够高效，能够在可用的操作时隙中运行，并且必须在100,000到100,000个处理器上很好地扩展。该项目的第一阶段（2011年2月至2013年1月）解决了支撑开发的几个基本科学问题，包括准均匀水平网格的选择、水平离散化的选择、传输方案的选择、时间积分方案的选择，以及项目的一些计算机科学方面的问题。在简化的二维流体系统（浅水方程）中测试和评估了几种候选方法，并确定了少量有前途的方法，用于第二阶段的进一步开发。该项目的第二阶段将以第一阶段的进展为基础，开发一个三维的、完全可压缩的动力核心。阶段2的工作大致分为三个工作包：*垂直方面。离散化的稳定性和准确性主要取决于垂直坐标的选择，预测的热力学变量的选择，以及变量相对于彼此的垂直位置（“交错”）。它还将取决于如何评估的细节，例如，压力梯度项，特别是在陡峭的山脉附近，以及垂直离散化如何与水平离散化耦合。根据目前的了解，将制定和测试候选方案。*代码设计和开发。三维动力核心的代码将基于精心设计的软件框架。数值离散化及其并行实现之间的接口将被优化，以便对前者的修改需要对后者的了解最少。软件框架将是高度灵活的，因此它可以很容易地适应未来动态核心的演变，如网格结构的变化。*测试。复杂的数值算法的行为很难从理论上预测，即使当单个组件被很好地理解和测试时。因此，至关重要的是尽早全面检验拟议的提法，并在必要时加以修订。早期测试将侧重于项目第一阶段产生的浅水配方，以及垂直配方的一维（柱）和二维（垂直切片）原型。一旦代码可用，三维配方的测试将立即开始。

Historically, major improvements in the accuracy of numerical weather forecasts and climate simulations have come from the increased resolution enabled by the exponential growth in computer power. In order to achieve further gains in accuracy through further increases in resolution, it will be necessary to exploit the massively parallel computer architectures that are becoming available. However, current state-of-the-art operational algorithms are not expected to perform well beyond a few thousand processors: the grid structure of the traditional latitude-longitude grid means that interprocessor communication eventually but inevitably becomes a bottleneck.The overall aim of the proposed project is to develop a new, three-dimensional, fully compressible dynamical core suitable for operational global and regional weather and climate prediction, as well as for research use, on massively parallel machines, and to demonstrate its accuracy, efficiency, and scalability. The accuracy should be comparable to that of existing state of the art algorithms. The algorithm must be efficient enough to run in the available operational time slots, and it must scale well on 100,000 to 1000,000 processors.Phase 1 of this project (Feb 2011 - Jan 2013) addressed several of the basic scientific questions that underpin the development, including choice of quasi-uniform horizontal grid, choice of horizontal discretization, choice of transport scheme, time integration scheme, and some of the computer science aspects of the project. Several candidate approaches were tested and evaluated in a simplified two-dimensional fluid system (the Shallow Water Equations), and a small number of promising approaches were identified for further development in Phase 2.Phase 2 of this project will build on the progress made in Phase 1 in order to develop a three-dimensional, fully compressible dynamical core. The work in Phase 2 falls broadly into three work packages:* Vertical aspects. The stability and accuracy of the discretization depends crucially on the choice of vertical coordinate, the choice of thermodynamic variables predicted, and the vertical placement of variables relative to each other (`staggering'). It will also depend on the details of how, for example, the pressure gradient term is evaluated, especially near steep mountains, and how the vertical discretization couples with the horizontal discretization. Building on current understanding, candidate schemes will be formulated and tested.* Code design and development. The code for the three-dimensional dynamical core will be based around a carefully designed software framework. The interface between the numerical discretization and its parallel implementation will be optimized, so that modifications to the former require minimal knowledge of the latter. The software framework will be highly flexible, so that it can easily accommodate future evolution of the dynamical core, such as changes in grid structure.* Testing. The behaviour of complex numerical algorithms can be difficult to predict theoretically, even when individual components are well understood and tested. It will be vital, therefore, to test comprehensively the proposed formulations at the earliest opportunity, and revise if necessary. Early testing will focus on the shallow water formulation arising out of Phase 1 of the project, and on one-dimensional (column) and two-dimensional (vertical slice) prototypes of the vertical formulation. Testing of the three-dimensional formulation will begin as soon as code is available.