Accelerating the FlowSimulator: Node-level performance analysis of DLR high-performance CFD Solvers using Likwid

Abstract

To achieve the best possible time to solution in High Performance Computing codes, node-level performance is a key factor. Thus, this performance analysis focuses on the two computational fluid dynamics (CFD) solvers CODA (Finite Volume) and Musubi (Lattice Boltzmann) on the DLR cluster CARO. We identify the relevant kernels of our software using profiling and tracing with Score-P and take a deeper look into their performance using the hardware performance monitoring tool LIKWID. Depending on the solver, different metrics are applied (e.g. roofline, CPI, cache misses, MLUP/s). Finally, we analyze the results in detail and derive possible optimization strategies for CODA and Musubi and summarize our experiences and best practices with the employed tools.