Copyrights to these papers may be held by the publishers. The download files are preprints. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.
Xiaoming Li, María J. Garzarán and David Padua (Proc. International Symposium on Code Generation and Optimization (CGO), pp. 111-124, 2004)
A Dynamically Tuned Sorting Library
Preprint (418 KB)
Published paper (link to publisher)
Empirical search is a strategy used during the installation oflibrary generators such as ATLAS, FFTW, and SPIRAL to identify the algorithm or the version of an algorithm that delivers thebest performance. In the past, empirical search has been appliedalmost exclusively to scientific problems. In this paper, we discuss the application of empirical search to sorting, which is oneof the best understood symbolic computing problems. When contrasted with the dense numerical computations of ATLAS, FFTW,and SPIRAL, sorting presents a new challenge, namely that the relative performance of the algorithms depend not only on the characteristics of the target machine and the size of the input data but also on the distribution of values in the input data set.Empirical search is applied in the study reported here as partof a sorting library generator. The resulting routines dynamicallyadapt to the characteristics of the input data by selecting the bestsorting algorithm from a small set of alternatives. To generate therun time selection mechanism our generator makes use of machinelearning to predict the best algorithm as a function of the characteristics of the input data set and the performance of the differentalgorithms on the target machine. This prediction is based on thedata obtained through empirical search at installation time.Our results show that our approach is quite effective. Whensorting data inputs of 12M keys with various standard deviations,our adaptive approach selected the best algorithm for all the inputdata sets and all platforms that we tried in our experiments. Thewrong decision could have introduced a performance degradationof up to 133%, with an average value of 44%.Keywords: Sorting, Numerical kernels we consider