Drishti

Drishti is a novel interactive web-based analysis framework to visualize I/O traces, highlight bottlenecks, and help understand the I/O behavior of scientific applications. Drishti aims to fill the gap between the trace collection, analysis, and tuning phases. The framework contains an interactive I/O trace analysis component for end-users to visually inspect their applications’ I/O behavior, focusing on areas of interest and getting a clear picture of common root causes of I/O performance bottlenecks. Based on the automatic detection of I/O performance bottlenecks, our framework maps numerous common and well-known bottlenecks and their solution recommendations that can be implemented by users.

Drishti Comparisons and Heatmaps

The proposed work will include investigating and building a solution to allow comparing and finding differences between two I/O trace files (similar to a diff), covering the analysis and visualization components. It will also explore additional metrics and counters such as Darshan heatmaps in the analysis and visualization components of the framework.

• Topics: I/O, HPC, data analysis, visualization, profiling, tracing
• Skills: Python, data analysis, performance profiling
• Difficulty: Moderate
• Size: Large (350 hours)
• Mentors: Jean Luca Bez and Suren Byna