The Learning and Inference Tutorial (LIT) for Large Parameter Spaces is a full day tutorial to be presented at ASPLOS 2008 on March 2, 2008 in Seattle, Washington (Web site). ASPLOS early registration (Web site) is available until February 11th, 2008. In the multidisciplinary spirit of ASPLOS, the learning and inference tutorial (LIT) provides practical insights into using statistical machine learning techniques that enable efficient, tractable data analysis for a broad range of domains. In recent years, these techniques have enabled fundamentally new capabilities in understanding relationships within parameter space exploration for circuits (delay and device sizes), microarchitectures (performance/power and computational resources), compilers (performance and optimization flags), and parallel software (performance and data decomposition). Inference and learning not only provide old answers to prior questions far more quickly, they also provide new answers to much larger, previously intractable questions. In particular, increasing system and algorithmic complexity combined with a growing number of tunable architectural parameters pose significant challenges for both simulator driven design evaluation and application performance modeling. In this hands-on tutorial the speakers present a series of robust techniques to address these challenges. The speakers first show how to apply statistical techniques such as clustering, association, and correlation analysis to understand application or architectural performance across large parameter spaces using sparse sampling. They then provide detailed instructions on how to construct two classes of effective predictive models based on piecewise polynomial regression and artificial neural networks. This tutorial will be a mixture of presentations, demonstrations, and hands-on exercises: every example used during the presentation will also be demonstrated live during the tutorial. In addition, attendees will have the chance to run all examples themselves during the hands-on sessions. This tutorial targets users with a need to understand performance (or other design metrics like power consumption) in large, high-dimensional parameter spaces or are interested in modeling the selected metric using just a small sampled subset of the complete design space. These techniques are helpful in scenarios with large spaces in which it is impossible or prohibitively expensive to simulate or compute the metric explicitly for each individual point in the space, including but not limited to architecture design studies, optimization problems, parameter tuning, and application optimization. The tutorial features a team of leading researchers in performance modeling from Harvard University, Cornell University and Lawrence Livermore National Laboratory. Presenters include: - David Brooks (Harvard University) - Bronis R. de Supinski (Lawrence Livermore Natl. Lab.) - Benjamin Lee (Harvard University) - Sally A. McKee (Cornell University) - Martin Schulz (Lawrence Livermore Natl. Lab.) - Karan Singh (Cornell University) The speakers have collectively performed fundamental research in statistical machine learning for the last few years in diverse application domains, including (1) performance modeling for parallel scientific computing kernels, (2) performance and power modeling for high-performance computer microarchitectures, (3) design optimization for heterogeneous and/or adaptive microarchitectures, and (4) comparative analysis of various sampling and predictive techniques in statistical inference and machine learning. This work led to several conference and journal publications. Eager to share their practical experiences, the speakers encourage others to leverage the computational efficiency of these techniques in learning and inference.