Pruning strategies in adaptive off-line tuning for optimized composition of components on heterogeneous systems

Abstract

We consolidate our convexity assumption that forms the basis for adaptive pruning of the sampling space.We provide better control of trade-offs between sampling time, runtime overhead and accuracy in adaptive empirical modeling.Reducing training time and improving prediction accuracy can be achieved simultaneously.Our method can converge faster and reaches higher accuracy than random sampling. Adaptive program optimizations, such as automatic selection of the expected fastest implementation variant for a computation component depending on hardware architecture and runtime context, are important especially for heterogeneous computing systems but require good performance models. Empirical performance models which require no or little human efforts show more practical feasibility if the sampling and training cost can be reduced to a reasonable level.In previous work we proposed an early version of adaptive sampling for efficient exploration and selection of training samples, which yields a decision-tree based method for representing, predicting and selecting the fastest implementation variants for given run-time call context's property values. For adaptive pruning we use a heuristic convexity assumption. In this paper we consolidate and improve the method by new pruning techniques to better support the convexity assumption and control the trade-off between sampling time, prediction accuracy and runtime prediction overhead. Our results show that the training time can be reduced by up to 39 times without noticeable prediction accuracy decrease.