Project Description

Physical Model Enhancement ­ Complex
chemical kinetics calculations enabled by GPUs
Hi-fidelity simulation of
combustors requires detailed chemical kinetic mechanisms that describe the
evolution of species concentration during chemical reactions. It is well
known that calculations for computing the chemical kinetic equations take
nearly 50-90% of the time in reactive CFD codes. These chemical kinetic
equations are ordinary differential equations (ODEs) and the mathematical
operations are all local in the computational mesh. In other words, these
calculations do not need communication with adjacent cores/processors and
hence are ideal for acceleration using hardware like Graphical Processing
Units (GPU). Therefore, UTRC is pursuing a novel approach for online
calculation of detailed chemical reactions using GPUs. UTRC¹s unique
algorithm enables efficient integration of numerically stiff ODE¹s and is
also ideally suited from the GPU memory standpoint. In stand-alone
calculations, this approach has shown significant reductions in
computational time (35X) compar
ed to running on a single core. Currently
we have implemented this approach in the LES solver LESLIE where all
operations except the chemistry computation will be performed on the CPU
cores and the chemistry will be computation will be performed solely on the
GPU. Our objective is to enable an arbitrary number of CPU cores
communicating with an arbitrary number of GPU to exploit multi-level
parallelism available in modern high-performance super-computers.

Numerical enhancements ­ Reduction of numerical error and subgrid-model
This will be a one-of-a-kind demonstration that for the same
physical model all reactive LES should converge to the same solution. It is
important to recognize that the simulations envisioned in (1) and (2) above
will become increasingly computationally expensive as the mesh is refined.
The unique HPC resources of Titan are expected to be critical in making this

Allocation History

Source Hours Start Date End Date