Perf Sampling Parser

⚠️ Do NOT use xperf / WPA / Windows Performance Toolkit to parse ETL files. Follow the steps below for ETL parsing.

Performance Analysis

Analyze per-process CPU usage from an ETL log; after the user selects target processes, export SpeedScope flame graphs.

Path Conventions

• <skill_dir>: directory containing this SKILL.md, derived from the path used to read this file
• Analysis scripts: <skill_dir>/scripts/ (tool paths are hard-coded inside the scripts; no external paths need to be passed in)

Inputs

• Required: path to the ETL file
• Required: confirm the PDB path with the user (semicolon-separated, NT symbol path format). If not provided, the script will automatically search the ETL's directory and _NT_SYMBOL_PATH.

Outputs

• Required: the output directory is fixed to the directory containing the ETL file.

Steps

Step 1: Analyze the process tree

python "<skill_dir>/scripts/process_tree.py" "<etl_path>"

Success marker: stdout contains DONE ... OK. On failure, read the first 50 lines of <etl_dir>/temp/pfv_log.txt and report them to the user.

Then read the generated JSON file (path comes from the ProcessTree written to ... line in stdout, typically <etl_dir>/<etl_basename_no_ext>.processtree.json).

Step 1.5: Parse the process tree JSON (using parse_processtree.py)

Use parse_processtree.py to parse the .processtree.json produced in Step 1 and present per-process CPU usage statistics:

python "<skill_dir>/scripts/parse_processtree.py" "<json_path>" --top 15

Sample output:

Total CPU time: 49630 ms

Process CPU Usage (Top 15):
16.18%  i4Tools  (pid=25344, 8030ms, parent=25188)
  7.63%  QtWebEngineProcess  (pid=19588, 3785ms, parent=25344)
11.71%  svchost  (pid=2516, 5812ms, parent=2036)
9.93%  Taskmgr  (pid=23840, 4927ms, parent=10268)
...

Total 442 processes with CPU samples

Process names for selection: i4Tools, QtWebEngineProcess, svchost, Taskmgr, ...

Optional arguments:

• --top N: show the top N processes (default 15)
• --min-cpu PCT: only show processes with CPU% >= PCT (default 0)
• --json-output: emit JSON for programmatic consumption

If no process has cpuPct > 0, tell the user "no CPU sampling data was captured" and list all processes for reference.

Step 2: Show the process list and let the user choose

Based on the Step 1.5 output, ask the user to choose 2-3 process names (without the .exe suffix) for analysis.

Step 3: Confirm the symbol path

After the user picks the processes, you must confirm the following two points with the user before proceeding:

1. Custom PDB path: ask whether there are extra PDB directories to add (local path or UNC, semicolon-separated). Leave empty if none.
2. Microsoft public symbol server: ask whether to include the MSFT public PDBs (SRV*C:\Temp\Symbols*https://msdl.microsoft.com/download/symbols). Adding it lets system DLL symbols resolve, but the first download is slow.

Example prompt:

Selected process: i4Tools

Please confirm the symbol configuration:
1. Any custom PDB path? (press Enter to skip)
2. Include the Microsoft public symbol server? (y/n — enables system DLL symbols but is slow on first run)

Assemble the final symbol path (NT symbol path format, semicolon-separated):

• Base: the ETL's directory (the script includes it automatically)
• Append the user's custom path (if any)
• If the user agrees, append SRV*C:\Temp\Symbols*https://msdl.microsoft.com/download/symbols

Example final <pdb_path>:

D:\myapp\pdbs;SRV*C:\Temp\Symbols*https://msdl.microsoft.com/download/symbols

Step 4: Export the SpeedScope JSON

Without PDBs (the user provided neither):

python "<skill_dir>/scripts/etl_to_speedscope.py" "<etl_path>" "<comma-separated process names>" "<etl_dir>"

With PDBs (the user supplied a custom path or opted into the MSFT public symbols):

python "<skill_dir>/scripts/etl_to_speedscope.py" "<etl_path>" "<comma-separated process names>" "<etl_dir>" "<pdb_path>"

Success marker: stdout contains DONE ... OK. On failure, read the first 50 lines of <etl_dir>/temp/pfv_log2.txt and report them to the user.

Step 5: Report the results

SpeedScope files generated (in the same directory as the ETL):
  \\server\share\log.processtree.json           (8030 ms CPU)
  \\server\share\QtWebEngineProcess_19588.speedscope.json (3785 ms CPU)

How to open: drag the file into https://speedscope.app (processed locally, not uploaded)

Symbol resolution failures (system DLLs or third-party modules without PDBs) are normal and do not affect application-level call stacks.

For details on PerfView command arguments and output format, read references/PerfViewCommands.md.