Overview

The project provides an easy way to save before/after CPU traces from Microbenchmark runs, and compare them visually using Differential Flame Graphs.

Areas where the code got slower are highlighted in red, while areas that are now faster are marked in blue; the intensity of the colour is proportional to the size of the difference.

See also the end-to-end demo (video).

Usage

Interacting with the script

• Overview of all commands: ./bench-flame-diff.sh -h
• Help for a specific command: ./bench-flame-diff.sh <command> -h

First usage

On first usage, initialise all dependencies by running: ./bench-flame-diff.sh init

General workflow

1. Run a specific Microbenchmark with CPU Stack sampling enabled (see below for instructions)
2. Save the trace as base for comparison using ./bench-flame-diff.sh save. It‘s worth picking a good names for the saved traces since you’re likely going to e.g. re-use the base while iterating on code changes.
3. Apply changes in your code and run the same benchmark as in step 1
4. Save the trace as current ./bench-flame-diff.sh save
5. Compare both traces using ./bench-flame-diff.sh diff which will create and open a diff in a web browser
6. Toggle between graphs using the buttons on the top:
   • base: flamegraph for the base trace
   • base-vs-curr: differential flame graph showing base vs current on the base trace
   • curr: flamegraph for the current trace
   • curr-vs-base: differential flame graph showing base vs current on the current trace
7. You can later go back to generated diffs using ./bench-flame-diff.sh open

Misc

Enabling stack sampling in Benchmark traces

This can be done in CLI or by editing build.gradle. Full documentation is here.

Quick CLI example:

# pick a target benchmark
tgt=:compose:foundation:foundation-benchmark:connectedCheck

# create a regex that targets a specific benchmark (test)
test_rx="androidx.compose.foundation.benchmark.lazy.LazyListScrollingBenchmark.scrollProgrammatically_noNewItems\[.*Row.*\]"

# run the benchmark and gather a 5 second (default) stack sample at 1000 Hz (default)
./gradlew $tgt -Pandroid.testInstrumentationRunnerArguments.tests_regex="$test_rx" \
 -P android.testInstrumentationRunnerArguments.androidx.benchmark.profiling.mode=StackSampling \
 -P android.testInstrumentationRunnerArguments.androidx.benchmark.profiling.sampleDurationSeconds=5 \
 -P android.testInstrumentationRunnerArguments.androidx.benchmark.profiling.sampleFrequency=1000

CLI completion

Generate shell-specific completion files with ./generate-completion.sh.

Then, source in your shell config, e.g.:

• For bash: dst="$(pwd)/completion_bash.sh"; echo "source '$dst'" >> ~/.bashrc
• For zsh: dst="$(pwd)/completion_zsh.sh"; echo "source '$dst'" >> ~/.zshrc

After restarting the shell session, you will be able to ‘tab-autocomplete’ commands and argument names.

Dependencies

On top of dependencies discoverable with ./gradlew app:dependencies the project depends on:

• https://github.com/brendangregg/FlameGraph
• https://android.googlesource.com/platform/system/extras/+/refs/heads/main/simpleperf/scripts

Both are fetched from the network in the init command and pinned to known-good-revisions.

Reporting issues

File an issue on Buganizer using this link or reach out directly to jgielzak@.

Known issues and future work items are tracked here.