Scalable Analysis Techniques for Microprocessor Performance Counter Metrics

摘要

Contemporary microprocessors provide a rich set of integrated performance counters that allow application developers and system architects alike the opportunity to gather important information about workload behaviors. These counters can capture instruction, memory, and operating system behaviors. Current techniques for analyzing data produced from these counters use raw counts, ratios, and visualization techniques to help users make decisions about their application source code. While these techniques are appropriate for analyzing data from one process, they do not scale easily to new levels demanded by contemporary computing systems. Indeed, the amount of data generated by these experiments is on the order of tens of thousands of data points. Furthermore, if users execute multiple experiments, then we add yet another dimension to this already knotty picture. This flood of multidimensional data can swamp efforts to harvest important ideas from these valuable counters. Very simply, this paper addresses these concerns by evaluating several multivariate statistical techniques on these datasets. We find that several techniques, such as statistical clustering, can automatically extract important features from this data. These derived results can, in turn, be feed directly back to an application developer, or used as input to a more comprehensive performance analysis environment, such as a visualization or an expert system.