GPU Bots & Pixel Wrangling
(December 2017: presentation on GPU bots and pixel wrangling: see slides.)
GPU Pixel Wrangling is the process of keeping various GPU bots green. On the GPU bots, tests run on physical hardware with real GPUs, not in VMs like the majority of the bots on the Chromium waterfall.
(Sorry, this link is Google internal only.)
These graphs show 1 day of activity by default. The drop-down boxes at the top allow viewing of longer durations.
See this CL for an example of how to update these graphs.
GPU Bots' Waterfalls
The waterfalls work much like any other; see the Tour of the Chromium Buildbot Waterfall for a more detailed explanation of how this is laid out. We have more subtle configurations because the GPU matters, not just the OS and release v. debug. Hence we have Windows Nvidia Release bots, Mac Intel Debug bots, and so on. The waterfalls we’re interested in are:
- Chromium GPU
- Various operating systems, configurations, GPUs, etc.
- Chromium GPU FYI
- These bots run less-standard configurations like Windows with AMD GPUs, Linux with Intel GPUs, etc.
- These bots build with top of tree ANGLE rather than the
- The ANGLE tryservers help ensure that these bots stay green. However, it is possible that due to ANGLE changes these bots may be red while the chromium.gpu bots are green.
- The ANGLE Wrangler is on-call to help resolve ANGLE-related breakage on this watefall.
- To determine if a different ANGLE revision was used between two builds, compare the
got_angle_revision buildbot property on the GPU builders or
parent_got_angle_revision on the testers. This revision can be used to do a
git log in the
The bots run several test suites. The majority of them have been migrated to the Telemetry harness, and are run within the full browser, in order to better test the code that is actually shipped. As of this writing, the tests included:
- Tests using the Telemetry harness:
- The WebGL conformance tests:
- A Google Maps test:
- Context loss tests:
- Depth capture tests:
- GPU process launch tests:
- Hardware acceleration validation tests:
- Pixel tests validating the end-to-end rendering pipeline:
- Stress tests of the screenshot functionality other tests use:
- drawElements tests (on the chromium.gpu.fyi waterfall): see
gles2_conform_test (requires internal sources): see
And more. See
src/testing/buildbot/README.md and the GPU sections of
waterfalls.pyl for the complete description of bots and tests.
Additionally, the Release bots run:
More details about the bots' setup can be found on the GPU Testing page.
- Ideally a wrangler should be a Chromium committer. If you're on the GPU pixel wrangling rotation, there will be an email notifying you of the upcoming shift, and a calendar appointment.
- If you aren‘t a committer, don’t panic. It's still best for everyone on the team to become acquainted with the procedures of maintaining the GPU bots.
- In this case you'll upload CLs to Gerrit to perform reverts (optionally using the new “Revert” button in the UI), and might consider using
TBR= to speed through trivial and urgent CLs. In general, try to send all CLs through the commit queue.
- Contact bajones, kainino, kbr, vmiura, zmo, or another member of the Chrome GPU team who's already a committer for help landing patches or reverts during your shift.
- Apply for access to the bots.
- You may want to install the Flake linker extension, which adds several useful features to the bot build log pages.
- Links to Chromium flakiness dashboard from build result pages, so you can see all failures for a single test across the fleet.
- Automatically hides green build steps so you can see the failure immediately.
- Turns build log links into deep links directly to the failure line in the log.
How to Keep the Bots Green
- Watch for redness on the tree.
- Sheriff-O-Matic now has support for the chromium.gpu.fyi waterfall!
- The chromium.gpu bots are covered under Sheriff-O-Matic's Chromium tab. As pixel wrangler, ignore any non-GPU test failures in this tab.
- The bots are expected to be green all the time. Flakiness on these bots is neither expected nor acceptable.
- If a bot goes consistently red, it‘s necessary to figure out whether a recent CL caused it, or whether it’s a problem with the bot or infrastructure.
- If it looks like a problem with the bot (deep problems like failing to check out the sources, the isolate server failing, etc.) notify the Chromium troopers and file a P1 bug with labels: Infra>Labs, Infra>Troopers and Internals>GPU>Testing. See the general tree sheriffing page for more details.
- Otherwise, examine the builds just before and after the redness was introduced. Look at the revisions in the builds before and after the failure was introduced.
- File a bug capturing the regression range and excerpts of any associated logs. Regressions should be marked P1. CC engineers who you think may be able to help triage the issue. Keep in mind that the logs on the bots expire after a few days, so make sure to add copies of relevant logs to the bug report.
- Use the
Hotlist=PixelWrangler label to mark bugs that require the pixel wrangler‘s attention, so it’s easy to find relevant bugs when handing off shifts.
- Study the regression range carefully. Use drover to revert any CLs which break the chromium.gpu bots. Use your judgment about chromium.gpu.fyi, since not all bots are covered by trybots. In the revert message, provide a clear description of what broke, links to failing builds, and excerpts of the failure logs, because the build logs expire after a few days.
- Make sure the bots are running jobs.
- Keep an eye on the console views of the various bots.
- Make sure the bots are all actively processing jobs. If they go offline for a long period of time, the “summary bubble” at the top may still be green, but the column in the console view will be gray.
- Email the Chromium troopers if you find a bot that's not processing jobs.
- Make sure the GPU try servers are in good health.
- The GPU try servers are no longer distinct bots on a separate waterfall, but instead run as part of the regular tryjobs on the Chromium waterfalls. The GPU tests run as part of the following tryservers' jobs:
[linux-rel] on the luci.chromium.try waterfall
[mac-rel] on the luci.chromium.try waterfall
[win7-rel] on the luci.chromium.try waterfall
- The best tool to use to quickly find flakiness on the tryservers is the new Chromium Try Flakes tool. Look for the names of GPU tests (like maps_pixel_test) as well as the test machines (e.g. mac-rel). If you see a flaky test, file a bug like this one. Also look for compile flakes that may indicate that a bot needs to be clobbered. Contact the Chromium sheriffs or troopers if so.
- Glance at these trybots from time to time and see if any GPU tests are failing frequently. Note that test failures are expected on these bots: individuals' patches may fail to apply, fail to compile, or break various tests. Look specifically for patterns in the failures. It isn't necessary to spend a lot of time investigating each individual failure. (Use the “Show: 200” link at the bottom of the page to see more history.)
- If the same set of tests are failing repeatedly, look at the individual runs. Examine the swarming results and see whether they‘re all running on the same machine. (This is the “Bot assigned to task” when clicking any of the test’s shards in the build logs.) If they are, something might be wrong with the hardware. Use the Swarming Server Stats tool to drill down into the specific builder.
- If you see the same test failing in a flaky manner across multiple machines and multiple CLs, it‘s crucial to investigate why it’s happening. crbug.com/395914 was one example of an innocent-looking Blink change which made it through the commit queue and introduced widespread flakiness in a range of GPU tests. The failures were also most visible on the try servers as opposed to the main waterfalls.
- Check if any pixel test failures are actual failures or need to be rebaselined.
- For a given build failing the pixel tests, look for either:
- One or more links named
gold_triage_link for <test name>. This will be the case if there are fewer than 10 links.
- A single link named
Too many artifacts produced to link individually, click for links. This will be the case if there are 10 or more links.
- In either case, follow the link(s) to the triage page for the image the failing test produced.
- Ensure you are signed in to the Gold server the links take you to (both @google.com and @chromium.org accounts work).
- Triage images on those pages (typically by approving them, but you can mark them as negative if it is an image that should not be produced). In the case of a negative image, a bug should be filed on crbug to investigate and fix the cause of that particular image being produced, as future occurrences of it will cause the test to fail. Such bugs should include the
Internals>GPU>Testing component and whatever component is suitable for the type of failing test (likely
Blink>Canvas). The test should also be marked as failing or skipped(see the item below on updating the Telemetry-based test expectations) so that the test failure doesn‘t show up as a builder failure. If the failure is consistent, prefer to skip instead of mark as failing so that the failure links don’t pile up. If the failure occurs on the trybots, include the change to the expectations in your CL.
- Additional, less common triage steps for the pixel tests can be found in this section of the GPU Gold documentation.
- Update Telemetry-based test expectations if necessary.
- Most of the GPU tests are run inside a full Chromium browser, launched by Telemetry, rather than a Gtest harness. The tests and their expectations are contained in src/content/test/gpu/gpu_tests/test_expectations . See for example webgl_conformance_expectations.txt, gpu_process_expectations.txt and pixel_expectations.txt.
- See the header of the file a list of modifiers to specify a bot configuration. It is possible to specify OS (down to a specific version, say, Windows 7 or Mountain Lion), GPU vendor (NVIDIA/AMD/Intel), and a specific GPU device.
- The key is to maintain the highest coverage: if you have to disable a test, disable it only on the specific configurations it's failing. Note that it is not possible to discern between Debug and Release configurations.
- Mark tests failing or skipped, which will suppress flaky failures, only as a last resort. It is only really necessary to suppress failures that are showing up on the GPU tryservers, since failing tests no longer close the Chromium tree.
- Please read the section on stamping out flakiness for motivation on how important it is to eliminate flakiness rather than hiding it.
- For the remaining Gtest-style tests, use the
DISABLED_ modifier to suppress any failures if necessary.
When Bots Misbehave (SSHing into a bot)
- See the Chrome Internal GPU Pixel Wrangling Instructions for information on ssh'ing in to the GPU bots.
Reproducing WebGL conformance test failures locally
- From the buildbot build output page, click on the failed shard to get to the swarming task page. Scroll to the bottom of the left panel for a command to run the task locally. This will automatically download the build and any other inputs needed.
- Alternatively, to run the test on a local build, pass the arguments
--browser=exact --browser-executable=/path/to/binary to
content/test/gpu/run_gpu_integration_test.py. Also see the telemetry documentation.
Extending the GPU Pixel Wrangling Rotation
See the Chrome Internal GPU Pixel Wrangling Instructions for information on extending the rotation.