| This dissertation proposes a new methodology for obtaining information about the relations emerging when compute-intensive applications are mapped onto advanced architectures. The proposed methodology incorporates knowledge and techniques from multiple areas that include statistics, operational research, pattern recognition, data mining, and performance evaluation to enable the extraction of performance information during the mapping process. The methodology is composed of four steps: problem analysis, design of experiments, data collection, and data analysis. In the first two steps, analyses of the application itself are completed to determine the appropriate design of experiments for establishing relations between changes is proposed for identifying important system metrics. An evaluation of different statistical analysis alternatives was carried out to characterize the types of data obtained in performance studies.; Several interesting results emerged from the application of this methodology on a computational electromagnetic case study. First, a correlation analysis embedded in the proposed methodology revealed that software instrumentation metrics exhibit collinearity. This implies a redundant information content in the data, limiting the set of statistical methods applicable for its analysis. Intrinsic dimensionality estimation and unsupervised feature subset selection identified the metrics containing the most performance information. On average, only 18% of the metrics were found to be important. Other results include identification of equivalency between multiple compiler options, reducing the actual set of options necessary at compile time. Also, a categorization of these options was obtained according to their effect in the application execution time. In summary, the application of the proposed methodology reveals that a detailed problem study preceding a systematic design of experiments, yields useful data on which appropriate statistical tools can provide unbiased information about the application-system interactions. Moreover the information obtained from this methodology can be converted into appropriate suggestions, observations, and guidelines for the scientific computing expert to tune applications to a particular computing system. (Abstract shortened by UMI.)... |
