Choosing the Best Kernel: Performance Models for Diffusion Operators in Particle Methods

Birte Schrader,S. Reboux,I. Sbalzarini

DOI: https://doi.org/10.1137/110835815

IF: 2.968

2012-06-19

SIAM Journal on Scientific Computing

Abstract:In scientific simulations of partial differential equations one is often faced with the task of choosing a discretization scheme or tuning the parameters of a discretized differential operator to perform well on a given problem. While this is mostly done through benchmark simulations on test problems, a problem-independent performance model would be desirable. Based on results from numerical analysis, we present a set of problem-independent performance measures for diffusion operators in particle methods. These measures quantify an operator's accuracy, stability, and computational cost. They can be explicitly derived in closed form, hence enabling comparisons between different operators and operator parameter tuning without the need for running any benchmark simulations. If a small number of benchmarks is available, a regression over the quality measures can be calibrated to absolute CPU time, hence defining predictive performance models for the different operators. We demonstrate this on the example of P...