perf-analyzer

Run the analyze_perf.py script with the output directory containing bubble_analysis.log as its argument. The script parses core metrics, computes utilization and bubble rate, and writes a performance_analysis_report.md into the same directory.

Bubble rate is the wait schedule time divided by the sum of AicoreTime and wait schedule time, expressed as a percentage. A bubble rate above 20% indicates heavy scheduling waits, often caused by small task granularity or poor scheduling strategy.

It is generated under output/output_<timestamp>/ relative to the working directory where the operator was executed with runtime debug mode enabled. If unsure, run find . -name bubble_analysis.log from the project root to locate it.

Core utilization equals AicoreTime divided by AicoreTime plus total wait time, where AicoreTime is total work time minus total wait time. Utilization above 90% earns a five-star rating; below 50% is rated one star.

The script fails when the operator was not run with runtime_debug_mode enabled or when the wrong directory is passed. Pass the timestamped output directory directly, or let the script recursively search subdirectories for bubble_analysis.log.

Follow the tuning direction section of the report: start with frontend tuning for low utilization, then swimlane-based deep tuning for high bubble rates, and finally in-core tuning. Each direction maps to a dedicated tuning skill such as tune-frontend or tune-swimlane.