Preparation of the NUMECA FINE/Open CFD Solver for Leadership Computing: Parallel I/O, Profiling, and Acceleration

Project Description

FINE/Open is a leader in high fidelity CFD offering solutions for a wide range of incompressible and compressible flows at low to hypersonic speeds, with many turbulence and fluid models available to the end user. Most industrial tools supporting a wide range of configurations in a user friendly package do so at the cost of HPC performance. Due to its modern programming model and efficient MPI implementation FINE/Open has the potential to be the first truly general use industrial CFD solver offering leadership scale parallelism with billion+ cell meshes. The implementation of parallel I/O will remove a known parallel bottleneck, and additional profiling and hybrid CPU/coprocessor execution will pave the way for leadership scale computing. The availability of an industrial CFD solver that achieves leadership scale parallel performance will be an asset to the business of NUMECA and the CFD industry on the whole.Additionally, high-order methods (HOM) based on Flux Reconstruction (FR) schemes are under development within FINE/Open. HOMs have gained a great popularity among the CFD community over the last decade. They have the potential of efficiently reducing numerical errors by resorting to an element-wise high-order polynomial representation of the solution, rather than by mesh refinement. The application of HOMs for LES/DNS simulations is very promising as turbulent scales can be better resolved by increasing the solution order, with fewer degrees of freedom than by reducing the element size to the desired resolution. HOMs based on FR schemes are well suited to run on massively parallelized systems. The stencil of the scheme involves only the adjacent elements sharing the same face, minimizing MPI communication. Furthermore, FR schemes require in general more computational power per degree of freedom than low order methods, and thus are tailored to run on GPU platforms or hybrid CPU-GPU clusters.