Accessible extreme-scale computing with Trixi.jl and the Julia programming language

使用 Trixi.jl 和 Julia 编程语言进行可访问的超大规模计算

• 批准号: 528753982
• 负责人: Professor Dr. Hendrik Ranocha
• 金额: --
• 依托单位: Center for Computational Engineering Science - Mathematics Division (MathCCES)
• 依托单位国家: 德国
• 项目类别: Research Grants
• 资助国家: 德国
• 项目状态: 未结题
• 来源: https://gepris.dfg.de/gepris/projekt/528753982?language=en
• 关键词: Accessible; extreme; scale; computing; Trixi.jl

Developing numerical simulation software for efficient high-performance computing (HPC) is difficult, especially when considering mesh adaptivity and heterogeneous hardware systems. A key issue is the need to balance the ability to add new functionality to the existing code base with the requirements of a scalable and robust implementation. The Julia programming language is uniquely positioned to bridge the gap between ease of use and efficiency in high-performance computing. However, a lot of this potential is still unlocked, since few HPC applications written in Julia exist and little documentation is available on best practices for Julia in the HPC context. A central aim of this project is therefore to extend the usability and impact of our Julia-based research software Trixi.jl by making extreme-scale computing as accessible as the rest of the Julia experience. Trixi.jl is a Julia package for adaptive numerical simulations of transport-dominated phenomena in science and engineering, with a focus on extensibility, ease of use, and high performance. To achieve our goal, we will first improve the modularization of the package and introduce new HPC-aware data structures that encapsulate some of the complexities of parallel data exchange and mesh adaptivity. A C-compatible application programming interface will then be introduced to allow Trixi.jl being used from other HPC simulation programs. To further ensure that the implementation retains a good code quality and high computational performance, we will integrate HPC hardware into our continuous testing infrastructure and create an HPC benchmarking suite for continuous performance monitoring. Furthermore, we will document all steps of the research software development process and publish best practices of Julia for HPC. Finally, we will promote the further development of Trixi.jl by providing tutorials and hands-on workshops for other scientists, and actively engage with the Julia HPC community.

开发用于高效高性能计算 (HPC) 的数值仿真软件很困难，特别是在考虑网格自适应性和异构硬件系统时。一个关键问题是需要平衡向现有代码库添加新功能的能力与可扩展且稳健的实现的要求。 Julia 编程语言具有独特的优势，可以弥合高性能计算的易用性和效率之间的差距。然而，这种潜力的很大一部分尚未被释放，因为用 Julia 编写的 HPC 应用程序很少，而且关于 Julia 在 HPC 环境中的最佳实践的文档也很少。因此，该项目的中心目标是通过使超大规模计算与 Julia 体验的其他部分一样易于访问，来扩展我们基于 Julia 的研究软件 Trixi.jl 的可用性和影响力。 Trixi.jl 是一个 Julia 软件包，用于对科学和工程中以传输为主的现象进行自适应数值模拟，重点关注可扩展性、易用性和高性能。为了实现我们的目标，我们将首先改进软件包的模块化，并引入新的 HPC 感知数据结构，封装并行数据交换和网格自适应的一些复杂性。然后将引入与 C 兼容的应用程序编程接口，以允许其他 HPC 模拟程序使用 Trixi.jl。为了进一步确保实施保持良好的代码质量和高计算性能，我们将把 HPC 硬件集成到我们的持续测试基础设施中，并创建一个用于持续性能监控的 HPC 基准测试套件。此外，我们将记录研究软件开发过程的所有步骤，并发布 Julia for HPC 的最佳实践。最后，我们将通过为其他科学家提供教程和实践研讨会来促进 Trixi.jl 的进一步发展，并积极与 Julia HPC 社区互动。