Perfcompare: Performance comparison try builder

The "perfcompare" try builder is an optional CQ try builder for measuring the performance impact of a change without landing it (i.e. for pre-submit performance testing). It runs performance tests both with and without a CL applied and compares their results to see if there were any performance regressions or improvements.

Googlers can refer to the Google-internal perfcompare docs for some additional documentation.

How to use it

For fuchsia.git CLs

To run perfcompare on a Gerrit CL, do the following:

  • Start a build: Select "Choose tryjobs" in the Gerrit Web UI, and select one or more of the perfcompare builders from the list of builders. A quick way to do that is to type "perfcompare" into the search field, which will filter the list to display the available perfcompare builders.
  • Get the results: A link to the try builder's results page will appear on the CL in Gerrit. When the builder run is finished, the results will be under "compare perf test results without and with CL" -> "stdout" (or "raw") on the build page.

These perfcompare builders are currently available and supported for running fuchsia.git's performance tests:

  • terminal.x64-release-perfcompare (recent builds): This runs fuchsia.git's performance tests on Intel NUCs (x64). This is the perfcompare version of the terminal.x64-release builder (i.e. it runs the same set of performance tests as that builder).

  • terminal.vim3-release-perfcompare (recent builds): This runs fuchsia.git's performance tests on VIM3s (ARM64). This is the perfcompare version of the terminal.vim3-release builder. Note that terminal.vim3-release is not run by the CQ by default, so it is more likely to be broken or have higher flake rates than other builders.

For integration.git CLs

Perfcompare is not supported yet for integration.git CLs.

Specifically, CLs that change dependencies in Jiri manifest files or jiri.lock files or that use patches.json are not yet supported by perfcompare. This includes CLs that change prebuilt packages, such as toolchain roll CLs.

Perfcompare does not know how to check out the source and prebuilt binaries before and after the CL in these cases, so it will give wrong results in these cases. It will produce a finding that there is no change in performance, even if the CL does change performance.

Example output

Here is part of the output from a perfcompare run on a simple test CL:

Summary counts:
  2939 test cases in total
  2938 test cases had no significant difference (no_sig_diff)
  1 test case got faster
  0 test cases got slower
  0 test cases added
  0 test cases removed

Results from test cases with differences:

Test case                                 Improve/regress?  Factor change  Mean before         Mean after
----------------------------------------  ----------------  -------------  ------------------  -----------------
fuchsia.microbenchmarks: ExampleNoOpLoop  faster            0.143-0.145    405.36 +/- 0.39 ns  58.49 +/- 0.30 ns

Results from all test cases:

Test case                                      Improve/regress?  Factor change  Mean before        Mean after
---------------------------------------------  ----------------  -------------  -----------------  -----------------
...
fuchsia.microbenchmarks: Syscall/ManyArgs      no_sig_diff       0.986-1.008    92.94 +/- 0.66 ns  92.65 +/- 0.40 ns
fuchsia.microbenchmarks: Syscall/Null          no_sig_diff       0.993-1.007    84.33 +/- 0.40 ns  84.31 +/- 0.19 ns
fuchsia.microbenchmarks: Thread/CreateAndJoin  no_sig_diff       0.950-1.034    34229 +/- 711 ns   33935 +/- 739 ns
fuchsia.microbenchmarks: TicksGet              no_sig_diff       0.981-1.022    19.77 +/- 0.19 ns  19.81 +/- 0.21 ns
...

Interpreting the results

  • no_sig_diff means that no statistically significant difference was found in the metric. It does not mean that there was no difference, just that any difference was too small (relative to the amount of variation in the metric) to be detected.

  • ci_too_wide is shown in the "Factor change" column if the confidence intervals in "Mean before" and "Mean after" are so wide that the lower bound went negative. This will happen for a metric if it has a large amount of variation.

Testing CL stacks versus individual CLs

The perfcompare builder measures the performance impact of individual CLs, not stacks of CLs.

As an example, suppose you have a series of CLs: P1, P2, P3, P4, P5, where P1 is the oldest (that is, all the other CLs depend on it). If you run perfcompare on P3, the "with CL" build will include P1+P2+P3, while the "without CL" build will include just P1+P2.

  • This provides a way to measure effects on test cases that haven't been landed yet. You can have one CL that adds a new performance test, and a follow-on CL that changes the software-under-test. Running perfcompare on the second CL will show how that CL affects the new test.
  • If you do want to measure the overall effect of a patch stack, one way to do that is to squash the changes into a single Git commit (such as with git merge --squash), upload that to Gerrit, and run perfcompare on that.

The "with CL" and "without CL" builds

The perfcompare builder applies the following steps sequentially to produce the "with CL" and "without CL" builds:

  1. Check out Fuchsia from the current tip-of-tree revision of integration.git.
  2. Apply the CL series to the checkout, up to and including the CL being tested. This uses jiri patch, which uses git rebase.
  3. Build Fuchsia. This gives the "with CL" build.
  4. Unapply the topmost CL from the checkout (leaving earlier CLs in the CL series, if any, applied). This works by running git checkout HEAD^ in the Git repo where the CL series was applied.
  5. Build Fuchsia again, doing an incremental build. This gives the "without CL" build.

Steps 1-3 are the same as for non-perfcompare Fuchsia try builders.

Limitations

  • CLs that use patches.json or that change dependencies in Jiri manifest files are not supported yet, as mentioned above.

How to run performance comparisons locally

The perfcompare builders use perfcompare.py to compare performance results. It is possible to use perfcompare.py to run performance tests locally (that is, not using Fuchsia Infra) and compare their results. See the documentation.

How to download the raw performance results

It is possible to download the raw performance test results produced by a perfcompare try builder run. This is useful if you want to modify the analysis that perfcompare.py performs. To do that, use the following steps:

  1. Find the values of the cas_instance and perfcompare_dataset_digest fields from the output properties of the perfcompare build. These can be found on the build page for the build (which is reachable from the "Checks" tab in the Gerrit code review). Examples of typical values are:

    • cas_instance="projects/chromium-swarm/instances/default_instance"
    • perfcompare_dataset_digest="3ff389154e02490f29e379564f7e70b3df66f74c3116ed50172cceec1e9d9888/165"

    For downloading results data from non-perfcompare builds, the field name to use is perf_dataset_digest rather than perfcompare_dataset_digest.

  2. Download the dataset by running the following command (using the prebuilt cas tool from the Fuchsia checkout):

    ./prebuilt/tools/cas/cas download -cas-instance $CAS_INSTANCE -digest $DIGEST -dir $DEST_DIR

  3. Run perfcompare.py on the downloaded dataset:

    python3 src/testing/perfcompare/perfcompare.py compare_perf $DEST_DIR/without_cl/ $DEST_DIR/with_cl/

Note that the RBE-CAS system keeps the data for only about 2-3 months, so the download command will fail if the build was not run recently. (The current default for the time-to-live (TTL) in RBE-CAS is 90 days.)