See Generating Profiles on Trybots and Using in Pinpoint.
See also go/chrome-pgo-internal(Googlers only and somewhat outdated).
Normally devs don't need to worry about this and can use the default profile for official builds. The default profile can be fetched by adding "checkout_pgo_profiles": True to custom_vars in the gclient config and running gclient runhooks.
To produce an executable built with a custom PGO profile:
Produce the instrumented executable using the following gn args:
chrome_pgo_phase = 1 enable_resource_allowlist_generation = false is_official_build = true symbol_level = 0 use_remoteexec = true
For android you need these in addition:
target_os = "android" target_cpu = "arm64"
Run representative benchmarks to produce profiles
python3 tools/pgo/generate_profile.py -C out/builddir
If collecting profiles on an android device, add a browser name like one of these:
python3 tools/pgo/generate_profile.py -C out/builddir \
--android-browser android-trichrome-chrome-google-bundle
You can find available browsers using:
tools/perf/run_benchmark run --browser=list
By default, some benchmark replay archives require special access permissions. For more details and to request access, please refer to Telemetry documentation. You can also choose to run generate_profile.py without these benchmarks, using the --run-public-benchmarks-only flag. However, note that doing so may produce a profile that isn't fully representative.
python3 tools/pgo/generate_profile.py -C out/builddir \ --android-browser android-trichrome-chrome-google-bundle \ --run-public-benchmarks-only
If generate_profile.py fails with ServiceException: 401 Anonymous caller does not have storage.objects.get access to the Google Cloud Storage object., then run download_from_google_storage --config (with your @google address; enter 0 as project-id).
This will produce out/builddir/profile.profdata
Produce the final PGO'd executable with the following gn args (and additional android args, if any):
enable_resource_allowlist_generation = false is_official_build = true symbol_level = 0 use_remoteexec = true pgo_data_path = "//out/builddir/profile.profdata"
To use a custom PGO profile, you need to specify the GCS bucket, path, and filename of the profile. This can be done by setting the following GN args:
pgo_gs_bucket = "chrome-pgo-trybot-profiles" pgo_gs_bucket_path = "pgo_profiles" pgo_override_filename = "<full-filename-of-your-profile>.profdata"
chrome_pgo_phase is defined in build/config/compiler/pgo/pgo.gni. This GN variable can be one of 0, 1, or 2, meaning “don't use profile”, “generating profile”, and “use profile” respectively. See pgo.gni for details on platform-specific GN variables that determine which phase is used in each build.
Which file under //chrome/build/pgo_profiles/ gets used? It depends on both the platform and _pgo_target. For example, for 64-bit android, the file //chrome/build/android-arm64.pgo.txt contains the name of the *.profdata file that is used as the PGO profile by default if no other profile is specified via the GN arg pgo_data_path.
On Android, the PGO profile generation process is tightly coupled with orderfile generation. The CI builders that generate PGO profiles are configured to trigger the corresponding orderfile generation builders upon successful completion.
This triggering mechanism is defined in the CI configuration files:
//internal/infra/config/subprojects/chrome/ci/chrome.pgo.star) specify which orderfile builders to trigger via the builders_to_trigger parameter.build_internal/recipes/recipes/clank/pgo.py reads this parameter and triggers the specified downstream builds.triggered_by parameter, which allows them to be triggered by the listed PGO builder.This ensures that the orderfile, which helps optimize the binary layout, is always built using the freshest profile data at the same Chromium commit that the PGO profile is generated from. The clank/pgo.py recipe also passes the Cloud Storage location of the newly generated PGO profile to the orderfile builder as it would not have been rolled into trunk by Skia autorollers at this point.
Step 1: pick an android-arm64-pgo build to benchmark against. In this example I picked the latest: https://ci.chromium.org/ui/p/chrome/builders/ci/android-arm64-pgo/12831/overview
Step 2: grab its chromium commit hash, your local CL should be rebased on it: 
In this case it is a26eb2940c446b477ffa908081083aefad361ead
Step 3: upload your change to gerrit and trigger the android-arm64-pgo trybot: 
See https://crrev.com/c/6979965?tab=checks
Step 4: find a successful run of android-arm64-pgo (you may need to do this several times), and copy the gsutil artifact name: 
For this one I got: chrome-android64-main-1759253260-88cce76848485c7544e4c3c3318a2d5ca1e4c3ec-2d4ffc91620fc1cae131415e81c8fed36dfabfaa.profdata (note that the second hash is not a26eb2940c446b477ffa908081083aefad361ead because I had a single modification CL on top of it).
Step 5: create two new CLs that sets this and the one generated by the original android-arm64-pgo CI bot in the default gn args file (IMPORTANT: these both need to be based on the same original base commit that generated them, in this case, both based on top of a26eb2940c446b477ffa908081083aefad361ead).
In this case, the one generated by CI was chrome-android64-main-1758734180-32cedafa6ae7a837a72fd6da923a5463d45bbc9a-a26eb2940c446b477ffa908081083aefad361ead.profdata
Trybot CL: https://crrev.com/c/7001460
CI CL: https://crrev.com/c/6998908 (note the different bucket for CI generated profiles chromium-optimization-profiles)
Step 6: Create a pinpoint job comparing the two above CLs: https://pinpoint-dot-chromeperf.appspot.com/ Example: 
This used android-pixel6-perf-pgo and 150 attempts to make results less noisy.
This is the actual pinpoint run that worked (some tweaks were needed): https://pinpoint-dot-chromeperf.appspot.com/job/118daa5a510000
Step 7: Analyze the pinpoint run as you would normally.
https://clang.llvm.org/docs/UsersManual.html#profile-guided-optimization