Comparisons of coupling strategies for massively parallel conjugate heat transfer with large eddy simulation

Abstract

The optimization of gas turbines is a complex multi-physic and multi-component problem that has long been based on engineer intuitions and expensive experiments or trial and error tests. Today, turbine experts commonly acknowledge that computer simulation is a very promising path for optimization, which can reduce costs and diminish the duration of the design process. Computations however remain a great challenge essentially because of the High Performance Computing (HPC) context, which is necessary for accurate estimates of real-life type of problems. Despite this difficulty, current highfidelity computer simulations become accessible for specific components of gas turbines [5]. These stand-alone simulations and solutions now face a new challenge: to improve the quality of the results, new physics must be introduced with specific and distinct numerical models. For example, in the context of multi-component simulations, further improving the accuracy of turbine wall temperature is of limited interest if wall temperature boundary conditions are still set approximately. Dealing with multi-physics, recent studies have shown interesting results by taking into account reactive flow as well as radiative and conductive heat transfers to predict wall temperature of a helicopter combustion chamber [2, 1]. Based on the simulation of conjugate heat transfer within an industrial combustor, the current study aims at comparing different strategies of code coupling on HPC architectures. The flow solver is the Large Eddy Simulation (LES) code AVBP already ported on massively parallel architectures [5]. The conduction solver is based on the same data structure and thus has the same performances in term of parallelism. Coupling these two codes although possible requires exchanging and treating information based on two different computational grids and time evolutions. Such transfers have to be thought to maintain code scalability while maintaining numerical accuracy, thus raising communication and HPC issues: transferring data from a distributed interface to an other distributed interface in a parallel way and on a very large number of processors is challenging and the solutions are not yet clear. The strategies investigated in this work go from standard client/server couplers to fully distributed couplers. Altough the standard client/server couplers are easier to implement, they appear to have scalability issues which fully distributed methods do not share.