profiling-analysis-profiling-main

Profiling Performance Analysis Main Skill

Feature Overview

This Skill is used to perform performance analysis on thestep_trace_time.csvfile generated by Profiling, automatically identify the system performance bottleneck type (computation, communication, dispatch), and trigger the corresponding sub-Skill for in-depth analysis.

Input Parameters

Parameter Name	Type	Required	Description
input_path	string	Yes	Input path: supports a singlestep_trace_time.csvfile path, or a folder path containing multiple such files

Usage Example

# 将路径修改为你的实际文件夹路径
TARGET_FOLDER = r"./"

Run Analysis

python scripts/performance_analysis_main_process.py

## 分析流程

1. **扫描CSV文件**：递归遍历指定目录，收集所有`step_trace_time.csv`文件
2. **解析数据**：验证文件结构，提取核心字段（Computing、Communication、Free）
3. **计算占比**：计算各阶段耗时占比
4. **瓶颈判定**：根据阈值判定瓶颈类型
5. **触发子Skill**：调用相应的子Skill进行深入分析

## 瓶颈判定规则

1. **下发问题（Hostbound）**：空闲耗时占比 > 10% → 触发 `/profiling-analysis-profiling-hostbound`
2. **计算问题（Computing）**：计算耗时占比 > 85% → 触发 `/profiling-analysis-profiling-computing`
3. **通信问题（Communication）**：通信耗时占比 > 10% → 触发 `/profiling-analysis-profiling-communication`

## 子Skill调用

如果遇到性能瓶颈问题，请调用子Agent运行对应的 /profiling-analysis-xxx-skill：

- **下发瓶颈问题**：调用 `/profiling-analysis-profiling-hostbound`
- **计算瓶颈问题**：调用 `/profiling-analysis-profiling-computing`
- **通信瓶颈问题**：调用 `/profiling-analysis-profiling-communication`