Computer simulations have become an integral part of the engine design and development process within the automotive industry. High-fidelity computational fluid dynamics (CFD) simulations of internal combustion (IC) engines enable the development and deployment of emission-reducing and fuel-saving technologies into the market faster than when these technologies are developed through experimental testing alone.

Within the automotive industry, simulations still rely heavily on lower-fidelity models, since these are relatively computationally inexpensive and afford turnaround times on the order of 12-24 hours, running on between 8 and 64 cores on standard Linux clusters. However, these simulations, though good for capturing overall trends, are not very accurate in terms of quantitative predictions of things such as emissions and cyclic variation, and engineers still rely significantly on costly and time-consuming experimental prototyping.

To move away completely from experimental testing, and further accelerate the engine design process, simulations need to be more predictive in nature. In order to achieve this, various levels of complexity need to be added to the computational models to account for things such as fuel chemistry effects, cycle to cycle stochastic variation in combustion, engine and injection system geometry effects, etc. To keep simulation turnaround time on the order of 12-24 hours while still incorporating higher fidelity models into CFD simulations, these simulations need to be scaled out significantly to on the order of thousands of cores.

This talk describes work done towards identifying and addressing several computational bottlenecks in the way of scaling a typical high-fidelity IC engine CFD simulation on several thousands of cores on Mira, towards enabling the use of supercomputers such as Mira to design fundamentally cleaner and more fuel efficient engines.