profile

name: profile description: Profile a benchmark workload with dotTrace sampling, generate an XML report, and analyze hot spots argument-hint: [--compare <baseline.xml>] [--filter ] [--callers ] [--top N]

dotTrace Profiling

Profile a Typhon benchmark workload using JetBrains dotTrace (sampling mode), export results via Reporter, and analyze with the Python script.

Prerequisites

These tools must be installed (one-time setup):

• dotnet tool install --global JetBrains.dotTrace.GlobalTools
• dotnet add package JetBrains.dotTrace.CommandLineTools.windows-x64 (for Reporter.exe)

Both are already installed in this project.

Tool Paths

DOTTRACE = dottrace (global tool, requires DOTNET_ROLL_FORWARD=LatestMajor on .NET 10+)
REPORTER = $USERPROFILE/.nuget/packages/jetbrains.dottrace.commandlinetools.windows-x64/2025.3.2/tools/Reporter.exe
ANALYZER = test/Typhon.Benchmark/profiling/analyze_profile.py
PROFILE_DIR = test/Typhon.Benchmark/profiling

If REPORTER path doesn't exist, look for the latest version:

ls "$USERPROFILE/.nuget/packages/jetbrains.dottrace.commandlinetools.windows-x64/"

Input

$ARGUMENTS may contain:

• <target> (required) — The benchmark target to profile. Can be:
  • A named profile target: --profile-delete (BTreeDeleteProfile)
  • A BDN filter: --filter *BTreeMicro*
  • A custom exe command to profile
• --compare <baseline.xml> — Compare current results against a baseline profile XML
• --filter <pattern> — Regex pattern to filter the analysis output (e.g., Remove, BTree)
• --callers <method> — Show callers of a specific method
• --top N — Number of top functions to show (default: 15)
• --save-as <name> — Save the snapshot/report with a descriptive name (default: btree-profile)

Workflow

Step 1: Build Release

cd test/Typhon.Benchmark && dotnet build -c Release

If build fails, report errors and stop.

Step 2: Profile with dotTrace

Determine the profiling command based on <target>:

For --profile-delete or other custom flags:

cd test/Typhon.Benchmark && DOTNET_ROLL_FORWARD=LatestMajor dottrace start \
  --profiling-type=Sampling \
  --save-to=profiling/<name>.dtp \
  --overwrite \
  --propagate-exit-code \
  -- bin/Release/net10.0/Typhon.Benchmark.exe <target-args>

For BDN --filter targets:

cd test/Typhon.Benchmark && DOTNET_ROLL_FORWARD=LatestMajor dottrace start \
  --profiling-type=Sampling \
  --save-to=profiling/<name>.dtp \
  --overwrite \
  --propagate-exit-code \
  -- bin/Release/net10.0/Typhon.Benchmark.exe --filter '<filter>'

IMPORTANT: Always profile the built executable directly (bin/Release/net10.0/Typhon.Benchmark.exe), NOT via dotnet run. Using dotnet run adds MSBuild noise to the profile.

IMPORTANT: Always use DOTNET_ROLL_FORWARD=LatestMajor — the dotTrace tool targets .NET 8 but we run .NET 10+.

Step 3: Generate XML Report

Use the pattern file to extract Typhon-only functions:

"$USERPROFILE/.nuget/packages/jetbrains.dottrace.commandlinetools.windows-x64/2025.3.2/tools/Reporter.exe" \
  report "profiling/<name>.dtp" \
  --pattern="profiling/btree-pattern.xml" \
  --save-to="profiling/<name>-report.xml" \
  --overwrite \
  --save-signature

The default pattern file (profiling/btree-pattern.xml) captures all Typhon\. methods. For other workloads, create a specific pattern file:

<Patterns>
  <Pattern PrintCallstacks="Full">Typhon\.</Pattern>
</Patterns>

Step 4: Analyze Results

Default analysis (full report):

cd test/Typhon.Benchmark && python profiling/analyze_profile.py profiling/<name>-report.xml

With filter:

python profiling/analyze_profile.py profiling/<name>-report.xml --filter "Remove" --top 20

Callers analysis:

python profiling/analyze_profile.py profiling/<name>-report.xml --callers "GetChunk"

JSON output (for programmatic use):

python profiling/analyze_profile.py profiling/<name>-report.xml --json

Step 5: Comparison (if --compare was specified)

If a baseline XML was provided, run the comparison:

python profiling/analyze_profile.py profiling/<baseline>.xml --compare profiling/<name>-report.xml

This shows per-function regressions and improvements between the two profiles.

Step 6: Display Summary

Present the results to the user:

1. Top 10 hot spots by self time — where CPU time is actually spent
2. Operation breakdown — time split by BTree operation (Add/Remove/TryGet)
3. Key observations — any surprising hot spots or optimization opportunities
4. If comparison mode: regressions and improvements

Example Usage

/profile --profile-delete
/profile --profile-delete --filter Remove
/profile --profile-delete --save-as before-opt
/profile --profile-delete --save-as after-opt --compare before-opt-report.xml
/profile --filter '*BTreeMicro*' --save-as btree-micro

Notes

• Snapshot files (.dtp) are ~50-100MB binary files — they are gitignored
• The XML report files are also gitignored (regenerated from snapshots)
• Pattern file (btree-pattern.xml) and analyzer script (analyze_profile.py) ARE tracked in git
• Sampling mode has ~5-10ms resolution — results are statistical approximations, not exact timings
• For finer granularity, use --profiling-type=Tracing instead of Sampling (much higher overhead)