The FeasNewt benchmark

Abstract

We describe the FeasNewt mesh-quality optimization benchmark. The performance of the code is dominated by three phases - gradient evaluation, Hessian evaluation and assembly, and sparse matrix-vector products - that have very different mixtures of floating-point operations and memory access patterns. The code includes an optional runtime data- and iteration-reordering phase, making it suitable for research on irregular memory access patterns. Mesh-quality optimization (or "mesh smoothing") is an important ingredient in the solution of nonlinear partial differential equations (PDEs) as well as an excellent surrogate for finite-element or finite-volume PDE solvers.