Implementation of Explicit Density-Based Unstructured CFD Solver for Turbomachinery Applications on Graphical Processing Units

Abstract

For the aerodynamic design of multistage compressors and turbines Computational Fluid Dynamics (CFD) plays a fundamental role. In fact it allows the characterization of the complex behaviour of turbomachinery components with high fidelity. Together with the availability of more and more powerful computing resources, current trends pursue the adoption of such high-fidelity tools and state-of-the-art technology even in the preliminary design phases. Within such a framework Graphical Processing Units (GPUS) yield further growth potential, allowing a significant reduction of CFD process turn-around times at relatively low costs. The target of the present work is to illustrate the design and implementation of an explicit density-based RANS coupled solver for the efficient and accurate numerical simulation of multi-dimensional time-dependent compressible fluid flows on polyhedral unstructured meshes. The solver has been developed within the object-oriented OpenFOAM framework, using OpenCL bindings to interface CPU and GPU and using MPI to interface multiple GPUS. The overall structure of the code, the numerical strategies adopted and the algorithms implemented are specifically designed in order to best exploit the huge computational peak power offered by modern GPUS, by minimizing memory transfers between CPUs and GPUS and potential branch divergence occurrences. This has a significant impact in terms of the speedup factor and is especially challenging within a polyhedral unstructured mesh framework. Specific tools for turbomachinery applications, such as Arbitrary Mesh Interface (AMI) and mixingplane (MP), are implemented within the GPU context. The credibility of the proposed CFD solver is assessed by tackling a number of benchmark test problems, including Rotor 67 axial compressor, C3X stator blade with conjugate heat transfer and Aachen multi-stage turbine. An average GPU speedup factor of approximately S ≈ 50 with respect to CPU is achieved (single precision, both GPU and CPU in 100 USD price range). Preliminary parallel scalability test run on multiple GPUS show a parallel efficiency factor of approximately E ≈ 75%

Comparison Of Acceleration Techniques on CFD Open-Source Software for Aerospace Applications

“…In particular, Residual Smoothing (RS), Local Time Stepping (LTS) and Multi-Grid (MG) are implemented and parallelized. 30 The solver is parallelized using OpenCL specifications. This means that two sets of source codes are written: one for the host (CPU) in C/C++ that make calls to OpenCL API functions and uses OpenCL data types, and one for the device (CPU, GPU or other accelerators) written in OpenCL C, a C99 based language with some restrictions.…”

A GPU-Accelerated Compressible RANS Solver for Fluid-Structure Interaction Simulations in Turbomachinery

“…The purpose of this section is to introduce the aerodynamic numerical formulations implemented in the solver. For a better description of the adopted formulations, the reader is referred to [9].…”

“…Besides the usual inflow, outflow and wall boundary conditions other types required for turbomachinery applications are implemented [9], such as Mixing Plane (MP) [23,24], AMI periodic boundary conditions, total and mass-flow inlet boundary conditions.…”

“…Since the explicit nature of the solver, convergence acceleration techniques like Multi-Grid (MG), Residual Smoothing (RS) and Local Time Stepping (LTS) [25,26,9] must be used to damp residuals and to obtain convergence rates that can be comparable with the ones provided by implicit solvers. LTS allows different cells to advance with different time-steps based on their geometrical and flow properties, accelerating in particular the convergence of the bigger cells far from the wall.…”

“…With the aim of an optimization process, another requirement of totally different nature arises: computational efficiency. This problem is typically addressed by implementing an efficient implicit formulation [7,8] or by accelerating explicit formulations with multiple combined convergence acceleration techniques [9]. Nowadays, GPUs provide a cheap and relatively easy way to obtain great SIMD (Single Instruction Multiple Data) floating point performance at the cost of a limited amount of available memory with respect to the usual amount of system RAM [10,9,11].…”