SC Conference - Activity Details

An Analysis Framework for Performance Data Mining and Knowledge-driven Performance Analysis

Kevin Huck  (University of Oregon)
Doctoral Research Showcase Session
Thursday,  04:30PM - 04:45PM
Room 17A/17B
Parallel performance diagnosis and engineering requires more sophisticated support for analysis automation, knowledge integration, intelligent problem discovery, and prescriptive feedback. How such capabilities are designed and implemented in a performance analysis framework includes consideration for programmability, extensibility, interoperability, and reuse. At the dawn of the petascale age, issues of scalability are also a serious concern both when analyzing the performance of thousands of cores and managing parametric studies with hundreds or thousands of experiments. Our approach to these complex analysis challenges is the design and development of PerfExplorer, a knowledge-engineering, rule-based, performance data mining framework. PerfExplorer provides dimension reduction, clustering, correlation and comparative analysis of parallel performance profiles. PerfExplorer also includes infrastructure for data persistence, provenance, and a scripting interface and inference engine to capture performance expert reasoning and build an analysis knowledge base for intelligent problem solving. PerfExplorer is distributed as part of the TAU performance system.
   IEEE Computer Society  /  ACM     2 0   Y E A R S   -   U N L E A S H I N G   T H E   P O W E R   O F   H P C