Towards Industrial Applicability of CFD Methodology for Massively-Parallel Hybrid Systems

This paper presents an overview of technical solutions aimed at improving robustness and industrial applicability of computational fluid dynamics (CFD) methodology for high-fidelity simulations on hybrid supercomputers. This methodology, incorporated in the NOISEtte code, is based on higher-accuracy numerical schemes on unstructured meshes and modern turbulence modelling approaches. When it comes to industrial applicability on modern supercomputer architectures, in addition to the basic theoretical properties of methods and models, such as approximation accuracy, many other factors arise. These factors can ruin the whole computational methodology if not properly addressed. For instance: compliance with modern restricted parallel paradigms; low memory consumption; robustness of the simulation algorithm, which implies applicability for a wide range of meshing tools and approaches; simplicity of numerical method configuration for the end user. Simulation algorithms must produce reasonable results in as wide range of application conditions as possible, often far from ideal. The present paper describes how a heterogeneous CFD code is being adapted to the cruel reality of practical applications: compliance with modern hybrid architectures, robustness of the numerical scheme and turbulence modeling approach, acceleration of solution, improving computing pattern and reducing memory consumption.