GPU Bot Details

This page describes in detail how the GPU bots are set up, which files affect their configuration, and how to both modify their behavior and add new bots.

Overview of the GPU bots' setup

Chromium‘s GPU bots, compared to the majority of the project’s test machines, are physical pieces of hardware. When end users run the Chrome browser, they are almost surely running it on a physical piece of hardware with a real graphics processor. There are some portions of the code base which simply can not be exercised by running the browser in a virtual machine, or on a software implementation of the underlying graphics libraries. The GPU bots were developed and deployed in order to cover these code paths, and avoid regressions that are otherwise inevitable in a project the size of the Chromium browser.

The GPU bots are utilized on the chromium.gpu and waterfalls, and various tryservers, as described in Using the GPU Bots.

All of the physical hardware for the bots lives in the Swarming pool, and most of it in the chromium.tests.gpu Swarming pool. The waterfall bots are simply virtual machines which spawn Swarming tasks with the appropriate tags to get them to run on the desired GPU and operating system type. So, for example, the Win10 x64 Release (NVIDIA) bot is actually a virtual machine which spawns all of its jobs with the Swarming parameters:

    "gpu": "10de:2184",
    "os": "Windows-10",
    "pool": "chromium.tests.gpu"

Since the GPUs in the Swarming pool are mostly homogeneous, this is sufficient to target the pool of Windows 10-like NVIDIA machines. (There are a few Windows 7-like NVIDIA bots in the pool, which necessitates the OS specifier.)

Details about the bots can be found on and by using src/tools/luci-go/swarming, for example swarming bots. If you are authenticated with credentials you will be able to make queries of the bots and see, for example, which GPUs are available.

The waterfall bots run tests on a single GPU type in order to make it easier to see regressions or flakiness that affect only a certain type of GPU.

The tryservers like win-rel which include GPU tests, on the other hand, run tests on more than one GPU type. As of this writing, the Windows tryservers ran tests on NVIDIA and Intel GPUs; the Mac tryservers ran tests on Intel and AMD GPUs. The way these tryservers' tests are specified is simply by mirroring how one or more waterfall bots work. This is an inherent property of the chromium_trybot recipe, which was designed to eliminate differences in behavior between the tryservers and waterfall bots. Since the tryservers mirror waterfall bots, if the waterfall bot is working, the tryserver must almost inherently be working as well.

There are some GPU configurations on the waterfall backed by only one machine, or a very small number of machines in the Swarming pool. A few examples are:

There are a couple of reasons to continue to support running tests on a specific machine: it might be too expensive to deploy the required multiple copies of said hardware, the hardware pool may have naturally died over time, or the configuration might not be reliable enough to begin scaling it up.

Adding a new isolated test to the bots

Adding a new test step to the bots requires that the test run via an isolate. Isolates describe both the binary and data dependencies of an executable, and are the underpinning of how the Swarming system works. See the LUCI documentation for background on Isolates and Swarming. Note that with the transition towards less Chromium-specific tools, you may see terms such as “CAS inputs” instead of “isolate”. These newer systems are functionally identical to the older ones from a user's perspective, so the terms can be safely interchanged.

Adding a new isolate

  1. Define your target using the template("test") template in src/testing/test.gni. See test("gl_tests") in src/gpu/ for an example. For a more complex example which invokes a series of scripts which finally launches the browser, see telemetry_gpu_integration_test in chrome/test/
  2. Add an entry to src/testing/buildbot/gn_isolate_map.pyl that refers to your target. Find a similar target to yours in order to determine the type. The type is referenced in src/tools/mb/

At this point you can build and upload your isolate to the isolate server.

See Isolated Testing for SWEs for the most up-to-date instructions. These instructions are a copy which show how to run an isolate that's been uploaded to the isolate server on your local machine rather than on Swarming.

If cd'd into src/:

  1. ./tools/mb/ isolate //out/Release [target name]
    • For example: ./tools/mb/ isolate //out/Release angle_end2end_tests
  2. ./tools/luci-go/isolate batcharchive -cas-instance chromium-swarm out/Release/[target name].isolated.gen.json
    • For example: ./tools/luci-go/isolate batcharchive -cas-instance chromium-swarm out/Release/angle_end2end_tests.isolated.gen.json See the section below on isolate server credentials.

Adding your new isolate to the tests that are run on the bots

See Adding new steps to the GPU bots for details on this process.

Relevant files that control the operation of the GPU bots

In the chromium/src workspace:

In the infradata/config workspace (Google internal only, sorry):

    • Defines a chromium.tests.gpu Swarming pool which contains all of the specialized hardware, except some hardware shared with Chromium: for example, the Windows and Linux NVIDIA bots, the Windows AMD bots, and the MacBook Pros with NVIDIA and AMD GPUs. New GPU hardware should be added to this pool.
    • Also defines the GCEs, Mac VMs and Mac machines used for CI builders on GPU and GPU.FYI waterfalls and trybots.
  • pools.cfg
    • Defines the Swarming pools for GCEs and Mac VMs used for manually triggered trybots.

Walkthroughs of various maintenance scenarios

This section describes various common scenarios that might arise when maintaining the GPU bots, and how they'd be addressed.

How to add a new test or an entire new step to the bots

This is described in Adding new tests to the GPU bots.

How to set up new virtual machine instances

The tests use virtual machines to build binaries and to trigger tests on physical hardware. VMs don't run any tests themselves. There are 3 types of bots:

  • Builders - these bots build test binaries, upload them to storage and trigger tester bots (see below). Builds must be done on the same OS on which the tests will run, except for Android tests, which are built on Linux.
  • Testers - these bots trigger tests to execute in Swarming and merge results from multiple shards. 2-core Linux GCEs are sufficient for this task.
  • Builder/testers - these are the combination of the above and have same OS constraints as builders. All trybots are of this type, while for CI bots it is optional.

The process is:

  1. Follow go/request-chrome-resources to get approval for the VMs. Use GPU project resource group. See this example ticket. You'll need to determine how many VMs are required, which OSes, how many cores and in which swarming pools they will be (see below for different scenarios).
    • If setting up a new GPU hardware pool, some VMs will also be needed for manual trybots, usually 2 VMs as of this writing.
    • Additional action is needed for Mac VMs, the GPU resource owner will assign the bug to Labs to deploy them. See this example ticket.
  2. Once GCE resource request is approved / Mac VMs are deployed, the VMs need to be added to the right Swarming pools in a CL in the infradata/config (Google internal) workspace.
    1. GCEs for Windows CI builders and builder/testers should be added to luci-chromium-gpu-ci-win10-8 group in
    2. GCEs for Linux and Android CI builders and builder/testers should be added to luci-chromium-gpu-ci-xenial-8 group in
    3. VMs for Mac CI builders and builder/testers should be added to builderfull_gpu_ci_bots group in Example.
    4. GCEs for CI testers for all OSes should be added to luci-chromium-gpu-ci-xenial-2 group in Example.
    5. GCEs and VMs for CQ and optional CQ GPU trybots for should be added to a corresponding gpu_try_bots group in Example. These trybots are “builderful”, i.e. these GCEs can't be shared among different bots. This is done in order to limit the number of concurrent builds on these bots (until is fixed) to prevent oversubscribing GPU hardware. win_optional_gpu_tests_rel is an exception, its GCEs come from luci-chromium-try-win10-*-8 groups in, see CL. This can cause oversubscription to Windows GPU hardware, however, Chrome Infra insisted on making this bot builderless due to frequent interruptions they get from limiting the number of concurrent builds on it, see discussion in CL.
    6. GCEs and VMs for manual GPU trybots should be added to a corresponding pool in “Manually-triggered GPU trybots” in If adding a new pool, it should also be added to pools.cfg. Example. This is a different mechanism to limit the load on GPU hardware, by having a small pool of GCEs which corresponds to some GPU hardware resource, and all trybots that target this GPU hardware compete for GCEs from this small pool.
    7. Run to regenerate configs/chromium-swarm/bots.cfg and configs/gce-provider/vms.cfg. Double-check your work there. Note that previously vms.cfg had to be edited manually. Part of the difficulty was in choosing a zone. This should soon no longer be necessary per, but consult with the Chrome Infra team to find out which of the zones has available capacity. This also can be checked on viceroy dashboard.
    8. Get this reviewed and landed. This step associates the VM or pool of VMs with the bot‘s name on the waterfall for “builderful” bots or increases swarmed pool capacity for “builderless” bots. Note: CR+1 is not sticky in this repo, so you’ll have to ping for re-review after every change, like rebase.

How to add a new tester bot to the waterfall

When deploying a new GPU configuration, it should be added to the waterfall first. The chromium.gpu waterfall should be reserved for those GPUs which are tested on the commit queue. (Some of the bots violate this rule – namely, the Debug bots – though we should strive to eliminate these differences.) Once the new configuration is ready to be fully deployed on tryservers, bots can be added to the chromium.gpu waterfall, and the tryservers changed to mirror them.

In order to add Release and Debug waterfall bots for a new configuration, experience has shown that at least 4 physical machines are needed in the swarming pool. The reason is that the tests all run in parallel on the Swarming cluster, so the load induced on the swarming bots is higher than it would be if the tests were run strictly serially.

With these prerequisites, these are the steps to add a new (swarmed) tester bot. (Actually, pair of bots -- Release and Debug. If deploying just one or the other, ignore the other configuration.) These instructions assume that you are reusing one of the existing builders, like GPU FYI Win Builder.

  1. Work with the Chrome Infrastructure Labs team to get the (minimum 4) physical machines added to the Swarming pool. Use or src/tools/luci-go/swarming bots to determine the PCI IDs of the GPUs in the bots. (These instructions will need to be updated for Android bots which don't have PCI buses.)

    1. Make sure to add these new machines to the chromium.tests.gpu Swarming pool by creating a CL against in the infradata/config (Google internal) workspace. Git configure your to if necessary. Here is one example CL and a second example.

    2. Run to regenerate configs/chromium-swarm/bots.cfg. Double-check your work there.

  2. Allocate new virtual machines for the bots as described in How to set up new virtual machine instances.

  3. Create a CL in the Chromium workspace which does the following. Here's an example CL.

    1. Adds the new machines to waterfalls.pyl directly or to mixins.pyl, referencing the new mixin in waterfalls.pyl.
      1. The swarming dimensions are crucial. These must match the GPU and OS type of the physical hardware in the Swarming pool. This is what causes the VMs to spawn their tests on the correct hardware. Make sure to use the chromium.tests.gpu pool, and that the new machines were specifically added to that pool.
      2. Make triply sure that there are no collisions between the new hardware you're adding and hardware already in the Swarming pool. For example, it used to be the case that all of the Windows NVIDIA bots ran the same OS version. Later, the Windows 8 flavor bots were added. In order to avoid accidentally running tests on Windows 8 when Windows 7 was intended, the OS in the swarming dimensions of the Win7 bots had to be changed from win to Windows-2008ServerR2-SP1 (the Win7-like flavor running in our data center). Similarly, the Win8 bots had to have a very precise OS description (Windows-2012ServerR2-SP0).
      3. If you‘re deploying a new bot that’s similar to another existing configuration, please search around in test_suite_exceptions.pyl for references to the other bot‘s name and see if your new bot needs to be added to any exclusion lists. For example, some of the tests don’t run on certain Win bots because of missing OpenGL extensions.
      4. Run to regenerate src/testing/buildbot/
    2. Updates or and its related generated files cr-buildbucket.cfg, luci-scheduler.cfg, and luci-milo.cfg:
      • Use the appropriate definition for the type of the bot being added, for example, ci.gpu_fyi_thin_tester() should be used for all CI tester bots on GPU FYI waterfall.
      • Make sure to set triggered_by property to the builder which triggers the testers (like 'GPU Win FYI Builder').
      • Include a ci.console_view_entry for the builder's console_view_entry argument. Look at the short names and categories to try and come up with a reasonable organization.
      • Make sure to set the serialize_tests property to True in the builder config. This is specified for waterfall bots but not trybots and helps avoid overloading the physical hardware. Additionally, if the bot is configured as a split builder/tester pair, ensure that the tester's builder config matches the parent builder and the tester is marked as being triggered by the parent builder.
    3. Run in src/infra/config to update the generated files. Double-check your work there.
    4. If you were adding a new builder, you would need to also add the new machine to src/tools/mb/mb_config.pyl.
  4. If the number of physical machines for the new bot permits, you should also add a manually-triggered trybot at the same time that the CI bot is added. This is described in How to add a new manually-triggered trybot.

While the above instructions assume that an existing parent builder will be be used, a new one can be set up by doing the same steps, but also adding the new parent builder at the same time. There are plenty of existing parent builder/child tester bots that you can use as a reference.

How to remove an existing bot from the waterfall

Basically, one needs to follow How to add a new tester bot to the waterfall step in reverse. To prevent bot failures during deletion process, pause the bot on

How to start running tests on a new GPU type on an existing try bot

Let's say that you want to cause the win-rel try bot to run tests on CoolNewGPUType in addition to the types it currently runs (as of this writing only NVIDIA). To do this:

  1. Make sure there is enough hardware capacity using the available tools to report utilization of the Swarming pool.
  2. Deploy Release and Debug testers on the chromium.gpu waterfall, following the instructions for the waterfall above. Make sure the flakiness on the new bots is comparable to existing chromium.gpu bots before proceeding.
  3. Create a CL in the chromium/src workspace that adds the new Release tester to win-rel's mirrors list. Rerun infra/config/
  4. Once the above CL lands, the commit queue will immediately start running tests on the CoolNewGPUType configuration. Be vigilant and make sure that tryjobs are green. If they are red for any reason, revert the CL and figure out offline what went wrong.

How to add a new manually-triggered trybot

Manually-triggered trybots are needed for investigating failures on a GPU type which doesn‘t have a corresponding CQ trybot (due to lack of GPU resources). Even for GPU types that have CQ trybots, it is convenient to have manually-triggered trybots as well, since the CQ trybot often runs on more than one GPU type, or some test suites which run on CI bot can be disabled on CQ trybot (when the CQ bot has no CI equivalent). Thus, all CI bots in chromium.gpu and have corresponding manually-triggered trybots, except a few which don’t have enough hardware to support it. A manually-triggered trybot should be added at the same time a CI bot is added.

Here are the steps to set up a new trybot which runs tests just on one particular GPU type. Let's consider that we are adding a manually-triggered trybot for the Win7 NVIDIA GPUs in Release mode. We will call the new bot gpu-fyi-try-win7-nvidia-rel-64.

  1. If there already exist some manually-triggered trybot which runs tests on the same group of machines (i.e. same GPU, OS and driver), the new trybot will have to share the VMs with it. Otherwise, create a new pool of VMs for the new hardware and allocate the VMs as described in How to set up new virtual machine instances, following the “Manually-triggered GPU trybots” instructions.

  2. Create a CL in the Chromium workspace which does the following. Here's a reference CL exemplifying the new “GCE pool per GPU hardware pool” way.

    1. Updates and its related generated file cr-buildbucket.cfg:
      • Add the new trybot with the right builder define and VMs pool. For gpu-fyi-try-win7-nvidia-rel-64 this would be gpu_win_builder() and luci.chromium.gpu.win7.nvidia.try.
      • Add the relevant CI bots to the new trybot's mirrors list. If the CI tester has a parent builder, the parent should be in the list as well.
    2. Run in src/infra/config to update the generated files. Double-check your work there.
    3. Adds the new trybot to src/tools/mb/mb_config.pyl and [src/tools/mb/mb_config_buckets.pyl][mb_config_buckets.pyl]. Use the same mixin as does the builder for the CI bot this trybot mirrors, in case of gpu-fyi-try-win7-nvidia-rel-64 this is GPU FYI Win x64 Builder and thus gpu_fyi_tests_release_trybot.
    4. Get this CL reviewed and landed.

At this point the new trybot should automatically show up in the “Choose tryjobs” pop-up in the Gerrit UI, under the luci.chromium.try heading, because it was deployed via LUCI. It should be possible to send a CL to it.

(It should not be necessary to modify buildbucket.config as is mentioned at the bottom of the “Choose tryjobs” pop-up. Contact the chrome-infra team if this doesn't work as expected.)

How to add a new try bot that runs a subset of tests or extra tests

Several projects (ANGLE, Dawn) run custom tests using the Chromium recipes. They use try bot bot configs that run subsets of Chromium or additional slower tests that can't be run on the main CQ.

These trybots are a little different because they do not mirror any waterfall bots.

Let's say the android_optional_gpu_tests_rel bot did not exist yet and you wanted to add it. The process is similar to adding a CI bot, but modifying slightly different files.

  1. Allocate new virtual machines for the bots as described in How to set up new virtual machine instances.
  2. Make sure there is enough hardware capacity using the available tools to report utilization of the Swarming pool.
  3. Create a CL in the Chromium workspace the does the following.
    1. Add your new bot android_optional_gpu_tests_rel to the waterfall in waterfalls.pyl. Here is an explicit example using the real bot.
    2. Re-run src/testing/buildbot/ to regenerate the JSON files.
    3. Add the bot definition to the relevant tryserver .star file. In this example, that would be Here is an explicit example using the real bot.
    4. Run in src/infra/config to update the generated files: luci-milo.cfg, luci-scheduler.cfg, cr-buildbucket.cfg. Double-check your work there.
    5. Update src/tools/mb/mb_config.pyl to include android_optional_gpu_tests_rel.
  4. After your CL lands you should be able to find and run android_optional_gpu_tests_rel on CLs using Choose Trybots in Gerrit.

How to test and deploy a driver and/or OS update

Let‘s say that you want to roll out an update to the graphics drivers or the OS on one of the configurations like the Linux NVIDIA bots. In order to verify that the new driver or OS won’t destabilize Chromium‘s commit queue, it’s necessary to run the new driver or OS on one of the waterfalls for a day or two to make sure the tests are reliably green before rolling out the driver or OS update. To do this:

  1. Make sure that all of the current Swarming jobs for this OS and GPU configuration are targeted at the “stable” version of the driver and the OS in waterfalls.pyl and mixins.pyl.

  2. File a Build Infrastructure bug, component Infra>Labs, to have ~4 of the physical machines already in the Swarming pool upgraded to the new version of the driver or the OS.

  3. If an “experimental” version of this bot doesn't yet exist, follow the instructions above for How to add a new tester bot to the waterfall to deploy one. However, you do not need to request additional GCE resources since there should be enough spare capacity in the GPU builderless pool to handle them. Additionally, ensure that the bot definition in includes a list_view argument specifying chromium.gpu.experimental.

  4. If an “experimental” version does already exist, re-add it to its default console in by uncommenting its console_view_entry argument and unpause it in the luci scheduler.

  5. Have this experimental bot target the new version of the driver or the OS in waterfalls.pyl and mixins.pyl. Sample CL.

  6. Hopefully, the new machine will pass the pixel tests. If it doesn‘t, then it’ll be necessary to follow the instructions on updating Gold baselines (step #4).

  7. Watch the new machine for a day or two to make sure it's stable.

  8. When it is, add the experimental driver/OS to the _stable mixin using the swarming OR operator |. For example:

    'win10_intel_hd_630_stable': {
      'swarming': {
        'dimensions': {
          'gpu': '8086:5912-|8086:5912-',
          'os': 'Windows-10',
          'pool': 'chromium.tests.gpu',

    This will cause tests triggered using the _stable mixin to run on either the old stable dimension or the experimental/new stable dimension.

    NOTE There is a hard cap of 8 combinations in swarming, so you can only use the OR operator in up to 3 dimensions if each dimension only has two options. More than two options per dimension is allowed as long as the total number of combinations is 8 or less.

  9. After it lands, ask the Chrome Infrastructure Labs team to roll out the driver update across all of the similarly configured bots in the swarming pool.

  10. If necessary, update pixel test expectations and remove the suppressions added above.

  11. Remove the old driver or OS version from the _stable mixin, leaving just the new stable version.

  12. Clean up the “experimental” version of the bot by pausing it in the luci scheduler and commenting out its console_view_entry argument in

Note that we leave the experimental bot in place. We could reclaim it, but it seems worthwhile to continuously test the “next” version of graphics drivers as well as the current stable ones.

Credentials for various servers

Working with the GPU bots requires credentials to various services: the isolate server, the swarming server, and cloud storage.

Isolate server credentials

To upload and download isolates you must first authenticate to the isolate server. From a Chromium checkout, run:

  • ./src/tools/luci-go/isolate login

This will open a web browser to complete the authentication flow. A email address is required in order to properly authenticate.

To test your authentication, find a hash for a recent isolate. Consult the instructions on Running Binaries from the Bots Locally to find a random hash from a target like gl_tests. Then run the following:

Swarming server credentials

The swarming server uses the same script as the isolate server. You will need to authenticate if you want to manually download the results of previous swarming jobs, trigger your own jobs, or run reproduce to re-run a remote job on your local workstation. Follow the instructions above, replacing the service with

Cloud storage credentials

Authentication to Google Cloud Storage is needed for a couple of reasons: uploading pixel test results to the cloud, and potentially uploading and downloading builds as well, at least in Debug mode. Use the copy of gsutil in depot_tools/third_party/gsutil/gsutil, and follow the Google Cloud Storage instructions to authenticate. You must use your email address and be a member of the Chrome GPU team in order to receive read-write access to the appropriate cloud storage buckets. Roughly:

  1. Run gsutil config
  2. Copy/paste the URL into your browser
  3. Log in with your account
  4. Allow the app to access the information it requests
  5. Copy-paste the resulting key back into your Terminal
  6. Press “enter” when prompted for a project-id (i.e., leave it empty)

At this point you should be able to write to the cloud storage bucket.

Navigate to to view the contents of the cloud storage bucket.