Function Overview
This Skill provides two independent analysis functions that can be selected according to user needs:
- Fast/Slow Card Performance Analysis: By parsing the analysis.db database file collected by MindStudio Insight, analyze the execution time and communication time of each NPU card, and identify fast and slow cards with inconsistent performance.
- Host-Side Trace Analysis: By parsing system trace files, analyze Host-side process performance, and identify problematic processes/threads and main bottlenecks.
The two functions run independently and do not call each other. Users can use the two analysis functions separately or simultaneously as needed.
Core functions:
- Multi-dimensional fast/slow card identification (computation time + communication time)
- Host-side process performance analysis (context switches, CPU migrations, interrupts, etc.)
- Detailed Excel report generation
- Problematic process/thread and bottleneck localization
Supported data formats
- Profiling data collected by MindStudio Insight (analysis.db database file)
- System trace files (such as trace files generated by Linux ftrace)
- Supports performance analysis of Ascend NPU systems
Sub-skill Description
This Skill contains two sub-skills, which users can select according to specific needs:
1. Fast/Slow Card Performance Analysis
Analyze in detail the performance differences between NPU cards, and identify fast and slow cards with abnormal computation or communication.
Main functions:
- Multi-dimensional fast/slow card identification based on computation time and communication time
- Calculation of performance difference percentage between cards
- Detailed Excel report generation
Detailed documentation: Fast/Slow Card Performance Analysis
2. Host-Side Trace Analysis
Analyze system trace files to identify Host-side process performance issues.
Main functions:
- Process runtime statistics
- Context switch count detection
- CPU migration count analysis
- Interrupt time statistics
- Problematic process/thread identification
Detailed documentation: Host-Side Trace Analysis
Usage
1. Automatically called by the main analysis Skill
This Skill is usually used by the main analysis Skill /profiling-analysis automatically triggered. When the main Skill detects that NPU inter-card performance differences or Host-side process bottlenecks need to be analyzed, it will automatically call the corresponding sub-skill for in-depth analysis.
2. Used separately
According to specific needs, select and use the corresponding sub-skill:
2.1 Using Fast/Slow Card Performance Analysis
python scripts/analyze_hostbound.py --db-path <analysis.db路径> [--output <报告输出路径>] [--speed-threshold <阈值>]
Detailed description: Please refer to Fast/Slow Card Performance Analysis Documentation
2.2 Using Host-Side Trace Analysis
python scripts/trace_analyzer.py --input <trace文件路径> [--output <报告输出路径>]
Detailed description: Please refer to Host-Side Trace Analysis Documentation
Usage Recommendations
Function Selection：
- When NPU inter-card performance differences need to be analyzed, use theFast/Slow Card Performance Analysissub-skill
- When Host-side process performance issues need to be analyzed, useHost-Side Trace AnalysisSub-skill
- To gain a comprehensive understanding of system performance, you can run both analysis sub-skills simultaneously
Analyze optimization paths：
- If inter-card performance imbalance is found, you can further analyze the Host-side Trace to identify possible process scheduling or resource contention issues
- If there is a performance bottleneck on the Host side, you can combine fast-slow card analysis to confirm whether it affects the overall performance of the NPU cards
Report interpretation：
- The fast-slow card analysis report focuses on inter-card performance differences
- The Host-side Trace analysis report focuses on process behavior and resource contention
- Combining the two reports provides more comprehensive system performance insights
Dependency requirements
- Python 3.8+
- Dependencies: pandas, openpyxl, matplotlib, sqlite3