Unstructured computational aerodynamics on many integrated core architecture

Optimizations of Unstructured Aerodynamics Computations for Many-core Architectures

“…Figure 6 presents the incremental progression of the flux kernel performance from the baseline to the most optimized code. The baseline code is the FORTRAN 77 code of [16] running out-of-the-box on KNL. The optimized code, on the other hand, is our reimplementation of PETSc-FUN3D that is written in C, and enhanced with several KNL-specific optimizations.…”

“…In the original PETSc-FUN3D code, the ψNKS phase is executed twice [11], [12], [13], [14], [16], [37] . The first one is a brief execution that performs a single pseudo time step with one nonlinear iteration to page in the data and warm up the memory hierarchy.…”

“…We relegate details of the modeling, discretization, numerical analysis, and solver to other publications [11], [14], [16], and focus herein upon the HPC and implementation aspects of the code.…”

“…This aims to preserve the spatial and temporal cache locality for the data items that are successively referenced in time; it is based upon the implementation of [11] with minor KNL-specific enhancements. Since this phase is primarily done through multiple vertex-based loops [13], [16], it is easy and straightforward to parallelize via OpenMP parallel for loop directives and static scheduling for the threads.…”

“…Later, It was extended in collaboration with Intel researchers to shared-memory optimizations and tuning on the multi-core Ivy Bridge Xeon architecture [14]. Recently, the code was ported into the first generation of Intel Xeon Phi architecture (Knights Corner (KNC) [15]), specifically exploring the IMCI instruction set of KNC in both modes, native and offload, as well as varying the thread distribution mechanisms across the cores [16].…”