Learning models in selfoptimizing systems

Library generators like ATLAS generate high-performance BLAS by performing a global, empirical search over the space of param-eter values. ATLAS generates programs based on the parameter values it finds, and evaluates the best one based on actual hard-ware performance. Yotov et al [8] replace the global search with an analytical model and get comparable performance. As such, the model-driven approach requires considerably less work to generate the BLAS. The tradeoff is that hand-built models require a deep understanding of both computer architectures and the algorithm in-volved, and can take years to conceive. We propose replacing the ATLAS search engine with a machine learning model. The advan-tages of this approach being: (1) unlike ATLAS, the machine learn-ing model requires minimal domain knowledge, (2) for one of our approaches, we obtain a model based on the parameter space in-stead of an optimal string of parameter values similar to Yotov et al [8]. Our approach can be extended to other library generators (e.g., FFTW) as well as new problems, with minimal effort. 1