Authors:
Thomas Rauber
1
and
Gudula Rünger
2
Affiliations:
1
Department of Computer Science, University of Bayreuth, Germany
;
2
Department of Computer Science, Chemnitz University of Technology, Germany
Keyword(s):
Numerical Solution Methods, Parameter Selection, Runtime Performance, Energy Consumption.
Abstract:
Compute-Bound numerical solution methods have a high demand for computational power and, thus, for energy. Both depend strongly on the numerical accuracy required for the approximation solution. A higher numerical accuracy often requires more execution time and energy. However, this dependence is more subtle and diverse. That means for a given numerical problem, different settings of the solution process, such as the use of different solvers, different implementation variants, different numbers of cores, or different operational frequencies result in a large number of different possibilities for the solution process, each of which may lead to a potentially different execution time and energy consumption. The best combination also depends on the specific execution platform used. Using different tolerance values for the time steps in the solution process adds another degree of complexity with a potentially different accuracy of the resulting approximation solution. The goal of this art
icle is to investigate the selection process of performance-optimal variants of all these computation possibilities when solving a given numerical problem. In particular, a selection process is proposed determining Pareto-optimal computation variants of the numerical method. As representative numerical solution method, explicit solution methods for ordinary differential equations are considered.
(More)