UC Davis

Performance visualization is an emerging field that adapts to the growing ecosystem of High-Performance Computing (HPC). With the continued growth in scale and complexity of HPC systems, code developers face the challenge of optimizing performance, requiring a detailed understanding of runtime behaviors of the system and the ability to identify and address performance bottlenecks. To meet this challenge, various tools have been developed to capture performance behaviors and conduct performance analysis, but the complexity and scale of the resulting data have made visualization systems essential for revealing and understanding key patterns that expose performance bottlenecks. The dissertation presents novel visual analytic methods that address the four key data challenges associated with performance analysis: Attribution, Scalability, Correlation, and Velocity. My contributions are driven by domain experts’ requirements to facilitate scalable interactive analysis of large-scale performance profiles and traces. Specifically, the dissertation addresses the challenges of analyzing and correlating collected performance data across three key domains of supercomputing — Hardware, Application, and Communication (HAC). Overall, the dissertation addresses the challenges of analyzing and correlating collected performance data in HAC domains, leading to a set of tools that are expected to greatly enhance HPC experts’ ability to understand and optimize the performance of demanding applications through interactive and scalable visual analytics.