The simulation of turbulent combustion is an extremely demanding computational problem. The complexity lies mainly in the nature of the combustion process itself, which is a multi-physics and multi-scale process with a high interaction of complex physical phenomena such as high-pressure fuel injection, atomization, vaporisation, fuel/air mixing, chemical reactions, turbulence/chemistry interaction, and pollutant formation.
From the numerical simulation perspective, this complexity translates into huge amounts of degrees of freedom to be tackled and, therefore, high computing requirements. This being said, can we consider that the accurate numerical simulation of realistic combustion problems represents an Exascale challenge?
There is no unique definition of an Exascale problem, but let’s put some numbers. We can estimate the computational time required for the type of simulations envisioned in the Center of Excellence for Combustion (CoEC) via extrapolation from the resources used in previous numerical studies. For instance, we can take as a reference the soot DNS simulation from Attili et al. (2014), considering heptane fuel at atmospheric pressure and a Reynolds number of 15000. That simulation employed 0.5 Billion points with 55 variables and 1 Billion Lagrangian tracers composed of 10 species each, corresponding to a total of about 40 Billion degrees of freedom. This study was performed on an entire IBM Blue Gene/P supercomputer with a peak performance of 222 Teraflops.
We could tackle the same problem but in more realistic conditions: a pressure of about 10-20 bar and a jet at Reynolds 100.000 where a realistic liquid fuel (an aviation fuel surrogate) is injected in the form of Lagrangian droplets. The five-fold increase in Reynolds number and the ten-fold increase in pressure requires an increase of the number of grid points in each direction of a factor of 10, i.e. a factor of 1000 in three dimensions. The use of a realistic liquid fuel requires a chemical mechanism with about 100 species, twice the number of the species in previous simulation. Taking into account factors such as the use of a Lagrangian point-particle approach for the liquid fuel droplets and the longer integration time required, it is possible to estimate a factor of about 10.000 for the computational cost of the envisioned simulation compared to the current state-of-the-art on combustion DNS. Assuming the same code efficiency with respect to the theoretical machine peak performance, a supercomputer with a peak performance of 222 Tera-flops x 10000 = 2 Exa-flops would be needed to perform a direct numerical simulation at engine conditions in reasonable time. With the use of large-eddy simulations (LES), where the smallest scales of the flow are modelled, the resolution requirements to characterise the turbulent flame characteristics can be relaxed and thus we can reduce the computing requirement or increase the geometrical details of the simulation domain.
Computational combustion is one of the disciplines that can benefit the most by the use of Exascale computing. Combustion simulations today cannot afford to achieve realistic engine conditions mainly due to the large computational requirements and the lack of efficient algorithms that can tackle high-pressure high-Reynolds number flames. The CoEC is tackling all the aspects of this challenge. Particularly, the project´s task on developing Exascale technologies (WP5) is focused on the efficient exploitation of (pre)Exacale supercomputers by the European combustion flagship codes that are presented in the consortium.
In the first months of the project, WP5 activity has focused on the analysis of the codes in collaboration with the Performance Optimization and Productivity (POP) Center of Excellence. The benchmarks considered cover various combustion simulation scenarios and will serve for both, defining the codes’ baseline performance and allowing a cross-comparison between them.
The activities of the work package are grouped in four main areas that will be developed along the project:
- Analysis an verification
- Development of scalable algorithms for combustion
- Exascale optimisations and performance portability
- Emerging technologies