ABSTRACT
The multiphysics, multiscale nature of turbulent chemically reacting flows makes these systems one of the most challenging to understand and control. In contrast to Reynolds-averaged Navier—Stokes, Large Eddy Simulation has historically employed spatial filtering techniques to split field variables into time-dependent resolved-scale and subfilter-scale components. Developing a complete understanding of relevant scales in the flow provides valuable information and enables optimal ways to resolve key features. A positive consequence of the approach is that the code is continually well positioned to achieve aggressive technical goals related to combustion science and the development of predictive models. As the compute performance is accelerated, the extent of time available to hide communication latencies becomes lower. This can expose previously hidden communication latencies that have a negative impact on parallel scalability. Programming models like OpenCL that promise to be portable across multiple vendors and architectures, have not been adopted widely and also suffer from lack of performance portability.
