D-Spline Performance Tuning Method Flexibly Responsive to Execution Time Perturbation

Various software automatic tuning methods have been proposed to search for the optimum parameter setting from among a combination of performance parameters. We have been studying a discrete spline (d-Spline)-based incremental performance parameter estimation (IPPE) method that does not require the approximation function to have differential continuity. In this method, a d-Spline generated from the minimum sample point is used to estimate the optimum value of the performance parameter. In prior methods, one measurement result was used to conduct sample point estimation; however, perturbations arising from the computing environment can affect estimates made in this manner. Such perturbations include disturbances introduced by the computing environment and OS jitters. In this study, we propose a method that considers execution time perturbation in performance parameter estimation by allowing for re-measurement under certain conditions by using an actual IPPE measurement. This lowers the inclusion of execution time perturbation in d-Spline approximation, thus enhancing the reliability of software automatic tuning.