SC Conference - Activity Details

Vijay S. Kumar  (Ohio State University)

Thursday,  04:45PM - 05:00PM

Room E145-146

Scientific applications that analyze terabyte-scale, multidimensional datasets are expressed as dataflow-graphs that are configured to accomodate a range of queries. End-users who cannot afford high execution times (due to various resource constraints) trade accuracy of analysis for performance gains, and supplement their queries with certain minimum quality requirements for the analysis to meet. My thesis focuses on performance optimizations for such applications, with emphasis on supporting accuracy-related trade-offs. The proposed framework allows users to express high-level constraints and quality requirements, and transforms these into appropriate low-level execution strategies for distributed computing environments. Using semantic representations of a domain's data and application characteristics, the framework allows explicit definition of various performance and accuracy-related parameters that influence application execution. Performance optimization is viewed as tuning of these parameters (by runtime systems) to determine their optimal values. Our framework has been successfully employed to support biomedical image analysis applications on multiple heterogeneous clusters.