Sustainable High-Performance Optimizations in SU2

Over a period of approximately 18 months, we have achieved an average 4-fold performance increase of the open source multiphysics suite SU2 through implementation optimizations (e.g. vectorization), and for some problems an additional 10-fold improvement via algorithmic changes. We have implemented a hybrid parallelization strategy (MPI + OpenMP) that improves the scalability of the code and allows key algorithms (such as multigrid) to maintain their effectiveness at small number of nodes per core. Our work has maintained the generality and versatility of the code by not relying on optimizations specific to given compilers, architectures, or physics. Furthermore, we maintain, or lower, the level of C++ knowledge needed for new developers. In this paper we document the implementation and algorithmic changes, give an overview of the details that allow implementing the hybrid parallel and vectorization frameworks in a way that hides the low-level complexity from high-level algorithm development. We demonstrate the improvements on benchmark problems known to the aeronautics community, and derive best practice guidelines to use the new capabilities.