 |
Center for Nonlinear Studies |
Power and cost constraints are driving future computer architectures to be radically different than the current generation of HPC architectures that have evolved iteratively over the last two decades, culminating in today's petascale computer architectures. We have proposed a Combustion Exascale Co-design Center (CECDC) to develop the necessary computational methodology to enable high-fidelity combustion simulations on these next-generation architectures. The overall project considers all aspects of the simulation process including hardware, programming models, data analysis and uncertainty quantification in addition to the core simulaton algorithms; however, this talk will focus primarily on the core PDE solution methodology. Although the details of future architectures are not known, we can anticipate some of their potential characteristics. In particular, future systems are likely to have heterogeneous, many-core nodes with reduced memory per core and increased costs of data movement relative to the cost of floating point operations. Numerical algorithms will need to expose high levels of concurrency within applications while accomodating these shifts in hardware characteristics. In this talk, we will review current practice for simulation of turbulent combustion with detailed chemistry and transport including basic discretization approaches, paralellization and current capabilities. We will then discuss how the architectural features discussed above impact the choice of algorithms and how these algorithms need to change to meet the challenge of emerging architectures. Finally, we will discuss some potential avenues of research to address these needs.
Host: Stephen Lee, srlee@lanl.gov, CCS -DO, see http://int.lanl.gov/orgs/ccs/csam