Automatic Software Performance Optimization on Modern Architectures

87 p.Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 2007.Frequent pattern mining is a fundamental problem in data mining and a large number of distinct algorithms have been proposed to solve it efficiently. However, no single algorithm outperforms all the others since their relative performance highly depends on the characteristics of the input data. In the dissertation, we present a machine learning based approach to select the best frequent pattern mining algorithm based on the input characteristics. Three of the fastest publicly available algorithms, FP_Growth, LCM and Eclat, were extensively evaluated using synthetic data sets. The results of these evaluations were used to train a support-vector machine (SVM) prediction system, which is then used at runtime to predict the best mining algorithm for real-world data sets. Our experiments show that the runtime prediction overhead is negligible and that the trained SVM prediction system usually identifies the best algorithm. In case of misprediction, the selected algorithm is still competitive in performance.U of I OnlyRestricted to the U of I community idenfinitely during batch ingest of legacy ETD