Acceleration of NOISEtte Code for Scale-Resolving Supercomputer Simulations of Turbulent Flows
- PDF / 1,081,074 Bytes
- 12 Pages / 612 x 792 pts (letter) Page_size
- 25 Downloads / 163 Views
Acceleration of NOISEtte Code for Scale-Resolving Supercomputer Simulations of Turbulent Flows A. V. Gorobets1* , P. A. Bakhvalov1** , A. P. Duben1*** , and P. V. Rodionov1**** (Submitted by E. E. Tyrtyshnikov) 1
Keldysh Institute of Applied Mathematics of Russian Academy of Sciences, Moscow, 125047 Russia Received February 28, 2020; revised April 5, 2020; accepted April 20, 2020
Abstract—The present work is devoted to accelerating the NOISEtte code and lowering its memory consumption. This code for scale-resolving supercomputer simulations of compressible turbulent flows is based on higher-accuracy methods for unstructured mixed-element meshes and hierarchical MPI + OpenMP parallelization for cluster systems with manycore processors. We demonstrate modifications of the underlying numerical method and its parallel implementation, which consist, in particular, in using a simplified approximation method for viscous fluxes and mixed floating-point precision. The modified version has been tested on several representative cases. The performance measurements and validation results are presented. DOI: 10.1134/S1995080220080077 Keywords and phrases: turbulent flows, scale-resolving simulation, viscous fluxes, aerodynamics, aeroacoustics, parallel CFD, MPI + OpenMP.
1. INTRODUCTION The numerical solution of computational fluid dynamics (CFD) problems usually imposes rather high computing demands, especially in the case of scale-resolving supercomputer simulations. Such simulations, among other things, allow obtaining unsteady aerodynamic and aeroacoustic properties, as well as more accurately predict integral characteristics (lift, drag). A fine enough spatial mesh, a sufficiently small time step and a long enough time integration period (for accumulation of flow statistics) are required to resolve dynamics of turbulent structures correctly. All this leads to high computational costs. The scientific computing community spends a lot of effort on accelerating and cheapening CFD simulations. The development of mathematical models and simulation approaches is aimed at obtaining accurate results on increasingly coarse meshes. For instance, hybrid approaches that efficiently combine Reynolds-averaged Navier–Stokes (RANS) and Large Eddy Simulation (LES) methods allow to significantly reduce mesh resolution requirements. Among such approaches, the non-zonal Detached Eddy Simulation (DES) methods are widely used, comprehensively investigated and validated. The most advanced versions of this method are formulated in [1, 2]. Another direction that leads to significant acceleration is the use of massively parallel computing accelerators, such as graphics processing units (GPUs). Representative examples of porting codes to the GPU architecture can be found, for instance, in [3, 4]. In [3], a rather mature and advanced GPU implementation for scale-resolving simulations using high-order schemes is compared with CPU implementations of well-known commercial CFD codes in terms of computing cost and accuracy. In [4], a GPU code for simulations of flows at high
Data Loading...
