Comparative Performance Modeling of Parallel Preconditioned Krylov Methods
Kanika Sood, Boyana Norris, Elizabeth R. Jessup · 2017
Many scientific and engineering computations rely on the scalable solution of large sparse linear systems. Preconditioned Krylov methods are widely used and offer many algorithmic choices whose performance varies depending on the characteristics of the linear system. In previous work, we have shown that the performance of different Krylov methods at small scales can be modeled using a small number of features based on structural and numerical properties of the input linear system. In this paper, we focus on comparing the scalability of parallel Krylov methods given different input properties without requiring extensive empirical measurements. We consider the PETSc implementations of Newton-Krylov methods to produce scalability rankings based on our new comparative modeling approach. The model-based ranking is validated by comparison with empirical results on a numerical simulation of driven fluid flow in a cavity.