Computational Advances in Modeling Opposed-Flow Diffusion Flames with Detailed Chemical Kinetics

Author(s):

Christopher P. Stone

Abstract:

The opposed-flow diffusion flame model, OPPDIF, has been refactored to improve its computational performance when using chemical kinetics mechanisms with hundreds or thousands of chemical species. A novel, block-tridiagonal matrix LU factorization method was formulated and used within the flame model’s nonlinear (Newton–Raphson) solver to directly solve the Jacobian-vector linear system. The new method has a theoretical run-time that is a factor of more than four faster than the original factorization method, based on the LAPACK banded matrix format, and requires only half the storage. This was achieved by maintaining a block-tridiagonal matrix structure even while performing (partial) row pivoting for numerical stability. Additionally, new methods for assembling the system’s Jacobian matrix and approximating its condition number were implemented and benchmarked using multicore (multithreaded) parallelism. In benchmark tests performed with 48 cores with two Intel Xeon Skylake CPUs and a large chemical kinetics mechanism with more than 1,300 species, the algorithm optimizations reduced the run-time by a factor of 5.7 over a previous parallel implementation. With a reduced, 600-species version of the mechanism, the run-time was reduced by a factor of 3.8. In addition to the net performance improvement, the parallel scalability of the OPPDIF model was improved by more than 100% on the large mechanism, allowing more efficient use of high-performance computing resources.