Resources

  • Domain Hints: Reference files for specific performance areas: CPU, Graphics, I/O, IPC, Memory, Power.
  • Perfetto SQL Reference: Reference guidelines for translating intents into valid queries are located in the SQL reference. You must read this reference and follow its Execution Protocol for all SQL generation.

Setup Phase (Mandatory)

  1. Initialize Scratchpad (Chain of Evidence):
    • Maintain your working memory in a local scratchpad file located in the exact same directory as the target trace file.
    • Name the file using the trace's filename appended with _analysis.md (e.g., [trace_filename]_analysis.md). Before creating it, check if a file with that name already exists by listing the directory's contents—to avoid biasing your investigation, DO NOT read the file's contents to check for its existence. If it does, append an incrementing version number (e.g., _v2.md, _v3.md) until you find an available filename. You MUST hardcode this exact filename in all subsequent tool calls.
    • Use this scratchpad STRICTLY to log verified facts: timestamps, slice names, thread IDs (utid/tid), and thread states.
    • DO NOT write preliminary hypotheses or premature conclusions in the scratchpad. It is a strict Chain of Evidence.
  2. Review SQL Reference: Read the SQL reference in references/sql.md and follow its Execution Protocol for all SQL generation. Do not guess schemas.
  3. Target Resolution: If the user's request is broad (e.g., "why is the app slow?") and doesn't specify a package name:
    • Execute a query to identify the active application: sql INCLUDE PERFETTO MODULE android.startup.startups; SELECT package FROM android_startups;
    • If multiple packages are returned, ask the user to choose one. Save the chosen package_name to your scratchpad.

Investigation Protocol

Follow this iterative loop until you have isolated the definitive root cause(s):

1. Formulate Hypothesis

  • Prioritization: Form hypotheses using information from: user prompt > built-in hints (references/hints_*.md) > general knowledge.
  • Source Attribution: Explicitly mention the source of your hypothesis (e.g., "Based on hints_io.md...").
  • Focus Constraint: Focus on the primary bottleneck. Avoid investigating deep into binder transactions unless the user explicitly asks for it or there is no other obvious bottleneck.
  • State Reasoning: Briefly state your reasoning based on previous findings before generating a new query.

2. Plan and Collect Data

  • Metrics First: Start with a high-level view using trace metrics before diving into custom SQL (e.g., ./trace_processor --run-metrics android_startup).
  • Broad to Narrow: Begin with broad queries using minimal filters. Favor fuzzy matching (e.g., GLOB '*abc*') over exact matching.
  • Overlapping Time: When filtering by time, you MUST check for events that overlap with the target time range (e.g., start1 < end2 AND start2 < end1) to ensure you don't miss slices that span across the boundaries.

3. Analyze and Drill Down (Depth-First)

  • Evidentiary Rigor: Do not draw conclusions without explicit data.
  • Wall Time vs. CPU Time: Do not assume a long-running slice is actively computing. You MUST query the thread_state table for the exact timestamp window of suspicious slices to verify if the thread was Running, Runnable (waiting for CPU), or Sleeping/Uninterruptible Sleep (blocked).
  • Follow Dependencies: If a thread is blocked/waiting, you MUST find what it is waiting for (Binder, Lock, I/O, etc.). Cross process boundaries if necessary. You cannot conclude an investigation on a waiting thread without identifying the blocker.

4. Exhaustive Investigation (Do Not Give Up Early)

  • Multiple Bottlenecks: Complex performance issues rarely have a single cause. Do NOT stop your investigation after finding the first anomaly. Even if you find a major bottleneck (e.g., emulator graphics lag), you MUST continue searching for other independent system-wide issues (e.g., lock contention, I/O stalls).
  • Global Verification: Periodically perform a system-wide query for the longest running slices (ORDER BY slice.dur DESC) and most frequent D-states to ensure your local investigation hasn't missed a massive, unrelated system stall.
  • Persist Through Dead Ends: If a hypothesis is disproven or a query returns empty, do not conclude. Pivot your focus, broaden your search constraints (fuzzy matching, wider time windows), and continue the mission.

Final Report

Only when you have followed the entire chain of dependencies to the root cause(s) AND confirmed through exhaustive search that no other major bottlenecks exist: 1. Summarize your findings detailing the verified chain of evidence. 2. Conclude with: "This concludes the trace analysis. You can review the full chain of evidence in [scratchpad_filename]. Let me know if you would like me to drill down into any of these specific threads, or if you'd like help drafting a bug report."