Parallelization strategies for an implicit Newton-based reactive flow solver

The solution of reactive flows using fully implicit methods on distributed memory machines is investigated in detail. Three different parallel implementations of Newton's method are described and tested on the solution of two-dimensional laminar axisymmetric coflow diffusion flames. Each implementation has different computational requirements, both in the amount of communication among the processes and in the computational overhead due to the calculation of physical quantities at the interfaces between subdomains. An effective trade-off is established between communications and calculations so that the most communication-intensive implementation results in computational speedup only if the network is sufficiently fast. Benchmark results are presented for a variety of chemical mechanisms, grid decomposition techniques, and hardware. Parallelization efficiencies of about 80% and speedups of 20-100 are reported for most test cases. The method developed here is well suited for complex chemistry problems with very large mechanisms; in particular, the numerical solution of a laminar axisymmetric JP-8/air coflow diffusion flame with a 222-species mechanism is made possible using this approach.