Abstracting the hardware: Assembly algorithms for numerical weather prediction on emerging massively parallel architectures

天气预报和气候模拟都是建立在求解大型方程组的基础上的，这些方程组被称为偏微分方程（pde），它们描述了大气中的气流。到目前为止，解决偏微分方程的软件，包括现有的气象局统一模型，都是由那些在建模和编程方面都是专家的人为特定种类的硬件编写的。然而，计算机硬件正在发生巨大而迅速的变化。新的大规模并行平台（如图形处理单元（gpu）和多核CPU系统）的出现需要非常不同的低级编程方法，更重要的是，需要针对不同硬件平台的不同方法。这也需要对这些平台的最佳优化策略进行新的研究。目前编写科学软件的方法无法满足工业和政府资金资源的需求，也无法跟上硬件平台的变化。该项目将使用一种新的方法来测试统一模型核心的现有和候选的新离散化，该方法将数值计算的单个工作项的规范与这些工作项在并行计算机上的共享方式分开。这将有助于决定应该如何编写新模型。特别是，它将使我们能够区分那些在高度并行性中不能很好地工作的算法和那些需要在非常并行的环境中工作的不同实现的算法。它还将使同一算法在不同平台的不同实现中进行测试，以建立算法，实现和平台的最佳组合。通过这种方式，它还将为新的统一模型提供一些未来的证明，因为反映新型并行硬件的底层更改将可能实现，而无需为每个新计算机平台重写整个模型。

Weather forecasting and climate simulation are both founded on solving large systems of equations, called partial differential equations or PDEs which describe the flows in the atmosphere. Hitherto, the software which solves PDEs, including the existing Met Office Unified Model, has been written for specific sorts of hardware by people who had to be experts both in the science they were modelling and in programming. However, large and rapid changes are now occurring in computer hardware. The emergence of new massively parallel platforms, such as graphical processing units (GPUs) and multicore CPU systems is requiring very different approaches to low-level programming and, importantly, different approaches for different hardware platforms. It also requires new research into the best optimisation strategies for these platforms. The current approach to writing scientific software will not deliver the changes that are required within the industrial and government funding resources available nor will it keep up with changing hardware platforms. This project will test the existing and candidate new discretisations for the Unified Model core by using a new approach which separates the specification of individual work items of numerical computation from how those work items are shared out over a parallel computer. This will enable decisions about how a new model should be written. In particular, it will enable us to distinguish between those algorithms which just don't work very well at high levels of parallelism and those which need to be implemented differently to work in very parallel contexts. It will also enable the same algorithm to be tested in different implementations for different platforms to establish the best combinations of algorithm, implementation and platform. In this way it will also provide some future-proofing to a new Unified Model as low-level changes to reflect new types of parallel hardware will be possible without rewriting the whole model for each new computer platform.