Telescopic Approach for Extreme-scale Parallel Mesh Generation for CFD Applications

摘要

The long term goal of this project is to address the following two challenges related to Extreme-scale Anisotropic Mesh Generation Environments: (1) Design a multi-layered algorithmic and software framework for 3D tetrahedral parallel mesh generation using state-of-the-art fully functional mesh generation codes like AFLR3 and a novel telescopic approach for current and emerging parallel architectures with several layers in network and memory hierarchy. The proposed approach explores concurrency at all hardware layers using abstractions at: (a) medium-grain level for many cores within a single chip and (b) coarse-grain level, i.e., sub-region and sub-domain level using proper error metric- and application-specific discrete data and continuous decomposition methods. The objective is to achieve and sustain a billion-way concurrency the next 15 years by (ⅰ) leveraging concurrency at different granularity levels and (ⅱ) carefully mapping work units to take advantage of the memory and network hierarchies. (2) Design a Parallel Runtime System for extreme-scale dynamic and irregular computations like mesh generation and refinement for CFD solvers. The runtime system will provide support for: (a) One-sided explicit message passing, (b) Global name-space, © multi-threaded programming model for inter-layer interactions, (d) automatic load balancing, (e) customizable data-movement and load-balancing, and (f) domain-specific energy-efficient race-to-halt, concurrency throttling, and component-level (core and memory) power scaling. Preliminary results from early stages of this work on parallel isotropic 3D Delaunay-based guaranteed quality mesh generation suggest that it is feasible to meet industries expectations and NASA's CFD vision for 2030~1.