Dr. Jacqueline H. Chen is a distinguished member of technical staff at the Combustion Research Facility, Sandia National Laboratory in Livermore. Her primary field of research is computational combustion, which relies on high-fidelity combustion simulations to develop accurate predictive combustion models, which will be used to design more fuel-efficient, cleaner-burning vehicles, planes and power plants in the future.
The 2016 ISC High Performance conference has invited Chen to keynote on Tuesday, June 21, on the topic of advancing the science of turbulent combustion using petascale and exascale simulations. The ISC Communications team caught up with Chen to find out more about combustion simulations and thus create more awareness for her research among a broad HPC audience.
ISC: What’s the thrust of your work and research at Sandia?
Jacqueline H. Chen: I am a computational combustion scientist at the Combustion Research Facility at Sandia. My work focuses on the development and application of a first principles direct numerical simulation approach to study fundamental ‘turbulence-chemistry’ interactions. The simulations are based on simple, laboratory configurations designed to isolate and elucidate underlying phenomena that may be present in real engines for transportation and power generation. These unit problems provide both new fundamental combustion science and validation data for the development of predictive models that will ultimately be used to design future fuel-efficient, clean engines.
I also lead a DOE ASCR sponsored Exascale Co-design Center, ExaCT (http://www.exactcodesign.org), a multi-disciplinary team of computer scientists, applied mathematicians and computational combustion scientists. The mission of ExaCT is to co-design all aspects of combustion simulation including numerical algorithms for partial differential equations, programming and execution models, scientific data management and analytics for in situ uncertainty quantification and graph-based topological analysis, and architectural simulations that explore hardware tradeoffs with combustion applications.
ISC: Can you give us a sense of how combustion simulation codes have impacted commercial engine and power plant designs thus far?
Chen: Recently, Cummins has used Reynolds-Averaged Navier-Stokes (RANS) models, which solves the time-averaged equations of motion for a fluid, to design heavy duty truck engines saving 10 to 15 percent in the development time and cost at the same time making the engine 10 percent more efficient.
In the future, industry will shift towards large-eddy simulation (LES), a more accurate and computationally intensive approach which resolves the energy-containing eddies and models turbulence and combustion at finer scales where energy and heat dissipate. LES will be used to capture cycle-to-cycle variability inherent in engines – which can lead to misfire for example — which RANS has difficulty capturing. Discovery and use-inspired computational research performed on the world’s largest supercomputers, in tandem with experiment and theory, is still needed, however, to develop predictive LES models in complex combustion regimes where future engines have to operate.
ISC: What will exascale systems do for combustion simulation codes that could not be achieved with petascale systems?
Chen: Exascale systems will enable fundamental high-fidelity combustion simulations capturing a larger dynamic range of turbulence scales, operating at higher pressure, and including a larger number of combustion compounds representative of large hydrocarbons and biofuels.
It will also enable more complex multi-physics including sprays, particulates and thermal radiation to be incorporated into these simulations. These high-fidelity simulations will be carefully designed to shed light on important underlying combustion science that is currently poorly understood and inspired by real applications. These particularly apply to low-temperature ignition processes in sprays coupled with turbulent mixing at high pressure or emissions characteristics in turbulent flames propagating into auto-igniting mixtures.
The massive data generated from these simulations, combined with experiments, will be used by scientists and engineers in academia and industry to develop and test new predictive models that work in more challenging combustion regimes, which future combustors will have to operate to realize gains in efficiency and to lower emissions.
ISC: Do you foresee a significant rewrite of legacy combustion simulation codes in order to take advantage of exascale machines? If so, who will end up doing that work?
Chen: Current petascale combustion simulation codes will have to be rewritten in order to take advantage of exascale machines. Current combustion simulation codes are written largely in a bulk synchronous programming approach which will not work at the exascale. Driven by power constraints, and the consequent challenges in resilience, and energy costs associated with data movement, exascale combustion codes will need to be rewritten. In response to these challenges, programming and execution models that tolerate asynchrony are needed along with new mathematical algorithms that minimize data movement and are inherently asynchronous.
Future predictive computational design tools for advanced combustion systems must be able to discern differences in physical and chemical properties of different fuels and couple that with the dynamic behavior of a combustor operating at high pressure and in highly turbulent environments. The numerical methodology needs to incorporate adaptive mesh refinement in the solution of large systems of partial differential equations with trillions of degrees of freedom to treat disparities in scales between flames and turbulence at high pressure. The core solver methodology is only one component of the required methodology. Disparity in growth rates of I/O systems and storage relative to compute throughput necessitate a full exascale workflow capability; current practice of archiving data for subsequent analysis will not be viable at the exascale. This full workflow also needs to support a wide range of in-situ analysis and uncertainty quantification methodologies.
The development of such a complex computational capability is most effectively achieved through combustion application co-design process involving an interdisciplinary team of computer scientists, applied mathematicians and computational combustion scientists. This team will work closely together to ensure that the future software stack, including new asynchrony-tolerant math algorithms for describing turbulent combustion, will work effectively on exascale hardware.
ISC: Will combustion codes have a major impact on co-design efforts? In particular, what hardware features are most important to these workloads?
Chen: Combustion codes have and continue to make a significant impact on co-design efforts across the entire stack — from mathematical algorithms for combustion simulation that reflect characteristics of future exascale architectures to asynchronous task-based programming and execution models that can adapt to node and system level non-uniformities, to numerous hardware features that support the end-to-end workflow of combustion simulations. Some of the hardware features identified through co-design that are most important to combustion workloads include larger register files, larger L1 caches for data reuse close to the processor core, fast interconnects for algebraic multigrid solvers used in low-Mach adaptive mesh refinement, software and hardware support for tasking-based programming models, and NVRAM and burst buffers to support complex and data-intensive interaction and data-exchange patterns, as well as managing data flow across complex storage hierarchies.