profiling-analysis

Ascend NPU Profiling Performance Analysis Skill Set

Function Overview

profiling-analysis is a complete Huawei Ascend NPU performance analysis skill set, used to analyze performance data generated by Profiling, automatically identify system performance bottleneck types (computing, communication, dispatch), and provide in-depth analysis capabilities and optimization suggestions.

Core Functions and Execution Workflow

The profiling-analysis skill automatically completes performance analysis through the following workflow:

1. File Scanning and Collection

Input: Target folder path
Processing: Recursively traverse the directory, automatically find all step_trace_time.csv files
Validation: Output an error message and terminate when no files are found

2. Performance Data Parsing and Calculation

Processing: Validate the file structure, extract core metrics such as Computing, Communication, and Free
Calculation: Calculate the time proportion of each stage (keep 2 decimal places)
Exception Handling: Skip files that fail to be read, have missing fields, or have a total duration of 0

3. Global Statistics and Bottleneck Determination

Statistics: Calculate the average time proportion of all files, and identify the file with the highest proportion for each item
Determination Logic(priority from high to low):
- Dispatch issue (Hostbound): Free time proportion > 20%
- Computing issue (Computing): Computing time proportion > 85%
- Communication issue (Communication): Communication time proportion > 10%
- No obvious bottleneck: all proportions are within the normal range

4. Automatic Invocation of Sub-skills

Based on the determined bottleneck type, automatically invoke the corresponding sub-skill for in-depth analysis:

Hostbound → Invoke profiling-analysis-hostbound sub-skill
Computing → Invoke profiling-analysis-computing sub-skill, execute the complete analysis workflow
Communication → Invoke profiling-analysis-communication sub-skill

5. Analysis Report Generation

Output a complete performance analysis report, including:

File scanning results
Single-file performance analysis data
Global statistics and the file with the highest proportion
Bottleneck determination conclusion
Sub-skill in-depth analysis results

Execution Workflow Diagram

┌─────────────────────────────────────────────────┐
│ 文件扫描与收集                                 │
│ 递归查找 step_trace_time.csv 文件               │
└─────────────────────────────────────────────────┘
                      ↓
┌─────────────────────────────────────────────────┐
│ 性能数据解析与计算                             │
│ 提取核心指标并计算占比                         │
└─────────────────────────────────────────────────┘
                      ↓
┌─────────────────────────────────────────────────┐
│ 全局统计与瓶颈判定                             │
│ 计算平均占比并识别瓶颈类型                     │
└─────────────────────────────────────────────────┘
                      ↓
┌─────────────────────────────────────────────────┐
│ 子技能自动调用                                 │
│ 根据瓶颈类型触发对应子技能                     │
└─────────────────────────────────────────────────┘
                      ↓
┌─────────────────────────────────────────────────┐
│ 分析报告生成                                   │
│ 输出完整性能分析结果                           │
└─────────────────────────────────────────────────┘

Input Parameters

Parameter Name	Type	Required	Description
input_path	string	Yes	Input path: supports a single Profiling file path, or a folder path containing multiple Profiling files

Usage Example

Invocation via AI tool (recommended)

调用 skill "profiling-analysis" 分析 "input_path" 中的性能数据

Parameter Description

input_path: Replace with the actual Profiling data folder path
This invocation will automatically complete the full workflow of file scanning, performance analysis, bottleneck determination, and sub-skill invocation

Independent Invocation of Sub-skills (optional)

If you need to directly call a specific sub-skill for in-depth analysis:

# 分析计算瓶颈
调用 skill "profiling-analysis-computing" 分析 "input_path" 中的高耗时算子

# 分析通信瓶颈
调用 skill "profiling-analysis-communication" 分析 "input_path" 中的通信性能

# 分析下发瓶颈
调用 skill "profiling-analysis-hostbound" 分析 "input_path" 中的Host侧问题

Sub-skill Introduction

profiling-analysis includes the following sub-skills, used for different types of performance analysis. When the main skill detects a corresponding bottleneck, it will automatically call the sub-skill for in-depth analysis, with no need for manual execution:

Sub-skill Name	Function Description	Applicable Scenario
profiling-analysis	Main analysis skill, identifies the type of performance bottleneck	First-time analysis of Profiling data, quickly locating the main bottleneck
profiling-analysis-computing(.\profiling-analysis-computing\SKILL.md)	Computation bottleneck analysis, including: - Screening of high-time-consuming operators (Top-N operators) - Operator performance data pivot table analysis - Shape parsing and optimization suggestions for key operators (such as the MatMul series)	When the computation proportion is too high, conduct in-depth analysis of the specific high-time-consuming operators and their performance characteristics
profiling-analysis-communication(.\profiling-analysis-communication\SKILL.md)	Communication bottleneck analysis, analyzing collective communication performance	When the communication proportion is too high, analyze the performance of communication operations
profiling-analysis-hostbound(.\profiling-analysis-hostbound\SKILL.md)	Dispatch bottleneck analysis, analyzing Host-side performance issues	When the idle proportion is too high, analyze Host-side dispatch issues

Supported Data Formats

step_trace_time.csv: Contains time proportion information such as computation, communication, and idle
op_statistic.csv: Contains operator execution statistics
kernel_details.csv: Contains detailed execution information of operator kernels
op_summary.csv: Contains operator execution summary information
analysis.db: Profiling database file

Installation and Dependencies

Python 3.8+
Dependencies: pandas, numpy, matplotlib (optional)

About this skill