Advanced Computation and I/O Methods for Earth-System Simulations (AIMES)

处理极端规模的地球系统模型是具有挑战性的。首先，模型代码不断被科学家修改和扩展，以纳入更多级别的细节。这增加了代码维护的负担，并限制了系统之间的性能移植。其次，随着科学家执行分辨率不断提高的运行或聚合集成运行的结果，所需的存储容量正在稳步增加。在AIMES项目中，我们的跨学科和国际团队致力于解决下一代二十面体地球系统模型中常见的可编程性、计算效率和I/O限制等关键问题。首先，我们将派生一个高级DSL作为表示，用于指定可以嵌入到现有代码中的这些模型的关键部分。这种高级表示允许科学家用更接近他们特定科学领域的抽象级别的术语来表达他们的代码，因此更自然。团队将从抽象概念中推导出Dynamico、ICON和NICAM模型的特定于模型的方言，并在原型中实现它们。其次，开发了一个概念和轻量级工具，用于将高级表示及其方言转换为各种现有的语言构造。通过选择适当的后端，可以创建适当的现有DSL和库方法，以及适当的内存布局和某些特定于体系结构的并行功能。虽然高级表示显著提高了可编程性，但科学模型跨系统的性能可移植性也得到了提高，两者都成为需要系统专家处理的正交方面。通过研究适用于二十面体数据的格式和改进有损压缩策略，改进了数据处理。在这个过程中，我们提出了克服当前限制的策略，实现了选定的原型，并与负责的联盟进行了沟通。为了展示开发的概念和组件的好处，将从所有模型中提取一个共享的开源基准套件，并对其进行小规模和大规模的评估。最终，我们打算通过在共享的想法和组件上进行合作来促进最佳实践和有用规范的开发，但我们也确保开发的工具和高级概念可以应用到其他领域。

Dealing with extreme scale Earth-system models is challenging. Firstly, model code is continuously revised and extended by scientists to incorporate further levels of detail.This increases the burden of code maintenance and limits performance portability across systems.Secondly, the required storage capacity is steadily increasing as scientists perform runs with growing resolution or aggregate results from ensemble runs. Within the AIMES project, our interdisciplinary and international team addresses the key issues of programmability, computational efficiency and I/O limitations that are common in next-generation Icosahedral earth-system models. Firstly, we will derive a high-level DSL as representation for specifying key parts of these models that can be embedded into existing code. This high-level representation allows scientists to express their code in a terminology that is closer to the abstraction level of their specific scientific domain and, thus, more natural.From the abstract concepts, the teams will derive model-specific dialects for the DYNAMICO, ICON and NICAM model and implement them in a prototype.Secondly, a concept and light-weight tool for translating the high-level representation and its dialects into a variety of existing language constructs is developed. By choosing the appropriate back-end, suitable existing DSLs and library approaches can be created together with an appropriate memory layout and certain architecture-specific features of parallelization. While the high-level representation increases programmability significantly, performance-portability of scientific models across systems is increased, and both become orthogonal aspects to be taken care of by system experts.Data handling is improved by investigating suitable formats for icosahedral data and by advancing lossy compression strategies. During this process, we propose strategies to overcome the current limitations, implement selected prototypes and communicate these with the responsible consortia.To demonstrate the benefit of the developed concepts and components, a shared open-source benchmark suite will be extracted from all models and evaluated on small and large scale.Ultimately, we intend to foster development of best-practices and useful norms by cooperating on shared ideas and components but we also ensure that developed tools and high-level concepts can be applied to other domains.