Mach Try Perf

To make it easier for developers to find the tests they need to run we built a perf-specific try selector called ./mach try perf. With this tool, you no longer need to remember the obfuscated platform and test names that you need to target for your tests. Instead, the new interface shows test categories along with a simplified name of the platform that they will run on.

When you trigger a try run from the perf selector, two try runs will be created. One with your changes, and one without. The push without your changes will be done on the revision that your patches are based on (which we call the base revision). In your console, after you trigger the try runs, you’ll find a PerfCompare link that will bring you directly to a comparison of the two pushes when they have completed.

The tool is built to be conservative about the number of tests to run, so if you are looking for something that is not listed, it’s likely hidden behind a flag found in the --help. Here’s a list of what you’ll find there:

$ ./mach try perf --help

optional arguments:
  -h, --help            show this help message and exit
perf arguments:
  --show-all            Show all available tasks. Alternatively, --full may be used.
  --android             Show android test categories (disabled by default).
  --chrome              Show tests available for Chrome-based browsers (disabled by
                        default).
  --custom-car          Show tests available for Custom Chromium-as-Release (disabled by
                        default). Use with --android flag to select Custom CaR android
                        tests (cstm-car-m)
  --safari              Show tests available for Safari (disabled by default).
  --safari-tp           Show tests available for Safari Technology Preview(disabled by
                        default).
  --live-sites          Run tasks with live sites (if possible). You can also use the
                        `live-sites` variant.
  --profile             Run tasks with profiling (if possible). You can also use the
                        `profiling` variant.
  --single-run          Run tasks without a comparison
  -q QUERY, --query QUERY
                        Query to run in either the perf-category selector, or the fuzzy
                        selector if --show-all/--full is provided.
  --browsertime-upload-apk BROWSERTIME_UPLOAD_APK
                        Path to an APK to upload. Note that this will replace the APK
                        installed in all Android Performance tests. If the Activity,
                        Binary Path, or Intents required change at all relative to the
                        existing GeckoView, and Fenix tasks, then you will need to make
                        fixes in the associated taskcluster files (e.g.
                        taskcluster/kinds/test/browsertime-mobile.yml). Alternatively, set
                        MOZ_FIREFOX_ANDROID_APK_OUTPUT to a path to an APK, and then run
                        the command with --browsertime-upload-apk firefox-android. This
                        option will only copy the APK for browsertime, see --mozperftest-
                        upload-apk to upload APKs for startup tests.
  --mozperftest-upload-apk MOZPERFTEST_UPLOAD_APK
                        See --browsertime-upload-apk. This option does the same thing
                        except it's for mozperftest tests such as the startup ones. Note
                        that those tests only exist through --show-all/--full as they aren't
                        contained in any existing categories.
  --detect-changes      Adds a task that detects performance changes using MWU.
  --comparator COMPARATOR
                        Either a path to a file to setup a custom comparison, or a builtin
                        name. See the Firefox source docs for mach try perf for examples
                        of how to build your own, along with the interface.
  --comparator-args [ARG=VALUE [ARG=VALUE ...]]
                        Arguments provided to the base, and new revision setup stages of
                        the comparator.
  --variants [ [ ...]]  Select variants to display in the selector from: fission,
                        bytecode-cached, live-sites, profiling, swr
  --platforms [ [ ...]]
                        Select specific platforms to target. Android only available with
                        --android. Available platforms: android-a51, android, windows,
                        linux, macosx, desktop
  --apps [ [ ...]]      Select specific applications to target from: firefox, chrome,
                        geckoview, fenix, chrome-m, safari, safari-tp, custom-car, cstm-
                        car-m
  --clear-cache         Deletes the try_perf_revision_cache file
  --alert ALERT         Run all tests that produced this alert summary ID based on the
                        alert summary table in either the alerts view or the regression
                        bug. The comparison that is produced will be based on the base
                        revision in your local repository (i.e. the base revision your
                        patches, if any, are based on). If only specific tests need to
                        run, use --tests to specify them (e.g. --tests webaudio).
  --extra-args [ [ ...]]
                        Set the extra args (e.x, --extra-args verbose post-startup-
                        delay=1)
  --non-pgo             Use opt/non-pgo builds instead of shippable/pgo builds. Setting
                        this flag will result in faster try runs.
  --tests [TESTS [TESTS ...]], -t [TESTS [TESTS ...]]
                        Select from all tasks that run these specific tests (e.g. amazon, or
                        speedometer3).

task configuration arguments:
  --artifact            Force artifact builds where possible.
  --no-artifact         Disable artifact builds even if being used locally.
  --browsertime         Use browsertime during Raptor tasks.
  --disable-pgo         Don't run PGO builds
  --env ENV             Set an environment variable, of the form FOO=BAR. Can
                        be passed in multiple times.
  --gecko-profile       Create and upload a gecko profile during talos/raptor
                        tasks.
  --gecko-profile-interval GECKO_PROFILE_INTERVAL
                        How frequently to take samples (ms)
  --gecko-profile-entries GECKO_PROFILE_ENTRIES
                        How many samples to take with the profiler
  --gecko-profile-features GECKO_PROFILE_FEATURES
                        Set the features enabled for the profiler.
  --gecko-profile-threads GECKO_PROFILE_THREADS
                        Comma-separated list of threads to sample.
  paths                 Run tasks containing tests under the specified
                        path(s).
  --rebuild [2-20]      Rebuild all selected tasks the specified number of
                        times.

Workflow

Below, you’ll find an overview of the features available in ./mach try perf. If you’d like to learn more about how to use this tool to enhance your developement process, see the Standard Workflow with Mach Try Perf page.

Standard Usage

To use mach try perf simply call ./mach try perf. This will open an interface for test selection like so:

Select the categories you’d like to run, hit enter, and wait for the tool to finish the pushes. Note that it can take some time to do both pushes, and you might not see logging for some time.

Retrigger

After the push is done, you will receive a Treeherder link that you can open to view your push. Access the Treeherder link to see all your tests.

To launch a retrigger, first select the task that you want to retrigger:

Then, click the rotating arrow icon in the task action bar, or press ‘r’ on your keyboard:

Additionally, you can add the flag --rebuild=2-20 to the try perf command to specify how many times you want to run the tests. If you want to learn more about retriggering please visit this page.

Add new jobs (mass retriggers)

The add new job function can be used to retrigger many tasks multiple times. To add a new job, follow these steps:

• Navigate to the push you want to add jobs on Treeherder.

• Click on the arrow drop-down on the top right of the push.

• Select the Custom push action from the menu.

You can copy the values from the target-tasks.json file from your Decision task and paste them into the task option. This method is useful for mass retriggers if needed. After you have pasted the json values, press the Trigger button.

Ideally, you should be able to use compare view to be more specific in the retriggers you do for tasks/tests that show a difference that they want to double-check.

Add extra-arguments/options to try run

To add additional arguments to a try tun, there are several approaches you can consider:

Use Treeherder

This method assumes that you already have the job that has been run and you want to run it again, but this time to add extra options as well. First select the task that you want to add extra options:

Then, click the three dots icon in the task action bar and select Custom Action:

A window will open where you need to select raptor-extra-options. There you can add all the options you need (e.g. extra_options: ‘verbose browser-cycles=3’). After finishing, press the Trigger button.

Modify the yml file

This method involves identifying the YML file that contains the test you are interested in and modifying or adding the extra-options key. Under this key you can add all the parameters you desire.

Use extra-args option

An alternative method is to utilize the --extra-args argument to try perf command (e.g. –extra-args verbose post-startup-delay=1).

Running Alert Tests

To run all the tests that triggered a given alert, use ./mach try perf --alert <ALERT-NUMBER>. Using this command will run all the tests that generated the alert summary ID provided in the regression bug. It’s recommended to use this when working with performance alerts. The alert number can be found in comment 0 on any alert bug such as this one. As seen in the image below, the alert number can be found just above the summary table. The comparison that is produced will be based on the base revision in your local repository (i.e. the base revision your patches, if any, are based on).

Running Tasks of a Specific Test

Using the --tests option, you can run all tasks that run a specific test. This is based on the test name that is used in the command that runs in the task. For raptor, this is the test specified by --test. For talos, it can either be a specific test in a suite like tp5n from xperf, or the suite xperf can be specified. For AWSY though, there are no specific tests that can be selected so the only option to select awsy tests is to specify awsy as the test.

If it’s used with --alert <NUM>, only the tasks that run the specific test will be run on try. If it’s used with --show-all or --full, you will only see the tasks that run the specific test in the fuzzy interface. Finally, if it’s used without either of those, then categories of the tests that were specified will be displayed in the fuzzy interface. For example, if --tests amazon is used, then categories like amazon linux firefox or amazon desktop will be displayed.

Chrome and Android

Android and chrome tests are disabled by default as they are often unneeded and waste our limited resources. If you need either of these, you can add --chrome and/or --android to the command like so ./mach try perf --android --chrome:

Variants

If you are looking for any variants (e.g. no-fission, bytecode-cached, live-sites), use the --variants options like so ./mach try perf --variants live-sites. This will select all possible categories that could have live-sites tests.

Note that it is expected that the offered categories have extra variants (such as bytecode-cached) as we are showing all possible combinations that can include live-sites.

Platforms

To target a particular platform you can use --platforms to only show categories with the given platforms.

Categories

In the future, this section will be populated dynamically. If you are wondering what the categories you selected will run, you can use --no-push to print out a list of tasks that will run like so:

$ ./mach try perf --no-push

Artifact builds enabled, pass --no-artifact to disable
Gathering tasks for Benchmarks desktop category
Executing queries: 'browsertime 'benchmark, !android 'shippable !-32 !clang, !live, !profil, !chrom
estimates: Runs 66 tasks (54 selected, 12 dependencies)
estimates: Total task duration 8:45:58
estimates: In the shortest 38% of durations (thanks!)
estimates: Should take about 1:04:58 (Finished around 2022-11-22 15:08)
Commit message:
Perf selections=Benchmarks desktop (queries='browsertime 'benchmark&!android 'shippable !-32 !clang&!live&!profil&!chrom)
Pushed via `mach try perf`
Calculated try_task_config.json:
{
    "env": {
        "TRY_SELECTOR": "fuzzy"
    },
    "tasks": [
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-firefox-ares6",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-firefox-assorted-dom",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-firefox-jetstream2",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-firefox-matrix-react-bench",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-firefox-motionmark-animometer",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-firefox-motionmark-htmlsuite",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-firefox-speedometer",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-firefox-stylebench",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-firefox-sunspider",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-firefox-twitch-animation",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-firefox-unity-webgl",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-firefox-webaudio",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-godot",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-godot-baseline",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-godot-optimizing",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-misc",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-misc-baseline",
        "test-linux1804-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-misc-optimizing",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-firefox-ares6",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-firefox-assorted-dom",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-firefox-jetstream2",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-firefox-matrix-react-bench",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-firefox-motionmark-animometer",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-firefox-motionmark-htmlsuite",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-firefox-speedometer",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-firefox-stylebench",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-firefox-sunspider",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-firefox-twitch-animation",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-firefox-unity-webgl",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-firefox-webaudio",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-godot",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-godot-baseline",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-godot-optimizing",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-misc",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-misc-baseline",
        "test-macosx1015-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-misc-optimizing",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-firefox-ares6",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-firefox-assorted-dom",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-firefox-jetstream2",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-firefox-matrix-react-bench",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-firefox-motionmark-animometer",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-firefox-motionmark-htmlsuite",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-firefox-speedometer",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-firefox-stylebench",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-firefox-sunspider",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-firefox-twitch-animation",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-firefox-unity-webgl",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-firefox-webaudio",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-godot",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-godot-baseline",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-godot-optimizing",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-misc",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-misc-baseline",
        "test-windows10-64-shippable-qr/opt-browsertime-benchmark-wasm-firefox-wasm-misc-optimizing"
    ],
    "use-artifact-builds": true,
    "version": 1
}

Adding a New Category

It’s very easy to add a new category if needed, and you can do so by modifying the PerfParser categories attribute here. The following is an example of a complex category that gives a good idea of what you have available:

"Resource Usage": {
    "query": {
        "talos": ["'talos 'xperf | 'tp5"],
        "raptor": ["'power 'osx"],
        "awsy": ["'awsy"],
    },
    "suites": ["talos", "raptor", "awsy"],
    "platform-restrictions": ["desktop"],
    "variant-restrictions": {
        "raptor": [],
        "talos": [],
    },
    "app-restrictions": {
        "raptor": ["firefox"],
        "talos": ["firefox"],
    },
    "tasks": [],
},

The following fields are available:

• query: Set the queries to use for each suite you need.

• suites: The suites that are needed for this category.

• tasks: A hard-coded list of tasks to select.

• platform-restrictions: The platforms that it can run on.

• app-restrictions: A list of apps that the category can run.

• variant-restrictions: A list of variants available for each suite.

Note that setting the App/Variant-Restriction fields should be used to restrict the available apps and variants, not expand them as the suites, apps, and platforms combined already provide the largest coverage. The restrictions should be used when you know certain things definitely won’t work, or will never be implemented for this category of tests. For instance, our Resource Usage tests only work on Firefox even though they may exist in Raptor which can run tests with Chrome.

Comparators

If the standard/default push-to-try comparison is not enough, you can build your own “comparator” that can setup the base, and new revisions. The default comparator BasePerfComparator runs the standard mach-try-perf comparison, and there also exists a custom comparator called BenchmarkComparator for running custom benchmark comparisons on try (using Github PR links).

If you’d like to add a custom comparator, you can either create it in a separate file and pass it in the --comparator, or add it to the tools/tryselect/selectors/perfselector/perfcomparators.py and use the name of the class as the --comparator argument (e.g. --comparator BenchmarkComparator). You can pass additional arguments to it using the --comparator-args option that accepts arguments in the format NAME=VALUE.

The custom comparator needs to be a subclass of BasePerfComparator, and optionally overrides its methods. See the comparators file for more information about the interface available. Here’s the general interface for it (subject to change), note that the @comparator decorator is required when making a builtin comparator:

@comparator
class BasePerfComparator:
    def __init__(self, vcs, compare_commit, current_revision_ref, comparator_args):
        """Initialize the standard/default settings for Comparators.

        :param vcs object: Used for updating the local repo.
        :param compare_commit str: The base revision found for the local repo.
        :param current_revision_ref str: The current revision of the local repo.
        :param comparator_args list: List of comparator args in the format NAME=VALUE.
        """

    def setup_base_revision(self, extra_args):
        """Setup the base try run/revision.

        The extra_args can be used to set additional
        arguments for Raptor (not available for other harnesses).

        :param extra_args list: A list of extra arguments to pass to the try tasks.
        """

    def teardown_base_revision(self):
        """Teardown the setup for the base revision."""

    def setup_new_revision(self, extra_args):
        """Setup the new try run/revision.

        Note that the extra_args are reset between the base, and new revision runs.

        :param extra_args list: A list of extra arguments to pass to the try tasks.
        """

    def teardown_new_revision(self):
        """Teardown the new run/revision setup."""

    def teardown(self):
        """Teardown for failures.

        This method can be used for ensuring that the repo is cleaned up
        when a failure is hit at any point in the process of doing the
        new/base revision setups, or the pushes to try.
        """

Frequently Asked Questions (FAQ)

If you have any questions which aren’t already answered below please reach out to us in the perftest matrix channel.

• How can I tell what a category or a set of selections will run?

  At the moment, you need to run your command with an additional option to see what will be run: ./mach try perf --no-push. See the Categories section for more information about this. In the future, we plan on having an dynamically updated list for the tasks in the Categories section of this document.

• What’s the difference between ``Pageload desktop``, and ``Pageload desktop firefox``?

  If you simply ran ./mach try perf with no additional options, then there is no difference. If you start adding additional browsers to the try run with commands like ./mach try perf --chrome, then Pageload desktop will select all tests available for ALL browsers available, and Pageload desktop firefox will only select Firefox tests. When --chrome is provided, you’ll also see a Pageload desktop chrome option.

• Help! I can’t find a test in any of the categories. What should I do?

  Use the option --show-all or --full. This will let you select tests from the ./mach try fuzzy --full interface directly instead of the categories. You will always be able to find your tests this way. Please be careful with your task selections though as it’s easy to run far too many tests in this way!

Future Work

The future work for this tool can be found in this bug. Feel free to file improvments, and bugs against it.