Folding Methods for Event Timelines in Performance Analysis

Abstract

The complexity of today's high performance computing systems, and their parallel software, requires performance analysis tools to fully understand application performance behavior. The visualization of event streams has proven to be a powerful approach for the detection of various types of performance problems. However, visualization of large numbers of process streams quickly hits the limits of available screen resolution. To alleviate this problem we propose folding strategies for event timelines that consider common questions during performance analysis. We demonstrate the effectiveness of our solution with code inefficiencies in two real-world applications, PIConGPU and COSMO-SPECS. Our methods facilitate visual scalability and provide powerful overviews of performance data at the same time. Furthermore, our folding strategies improve GPU stream visualization and allow easy evaluation of the GPU device utilization.