Contact: nduca, ernstm
Chrome now has an awesome rendering benchmark system for GPU and rendering related benchmarks. It works on all chrome flavors, even android and CrOS, even in their content_shell forms.
We use the following terminology:
- action: an interaction with a web page that we want to measure. e.g. scrolling, or simply waiting for a few seconds
- page set: a collection of web pages and associated actions
- metric: computes high-level statistical measures (e.g. means and medians) from raw data (e.g. traces).
- measurement: loads the pages from a page set, executes the associated actions, collects raw data, and computes final results using a metric.
- benchmark: bundles together a measurement and a page_set
To run the rendering benchmarks you need:
Once you've got these things, you're ready to go. To run our top 25 page set through our smoothness measurement (which tests scrolling speed for sites that scroll, or interaction speed for sites that have interactions):
./run_benchmark --browser=canary smoothness.top_25
To measure impl-side painting on important mobile sites:
tools/perf/run_benchmark --browser=canary smoothness.key_mobile_sites
To measure the key mobile sites on an attached android device:
tools/perf/run_benchmark --browser=android-chrome smoothness.top_25
To run the smoothness measurement on a Chrome OS device with IP address $CHROMEBOOK_IP from a host machine with a chromium checkout, do this:
--browser=cros-chromeos --remote=$CHROMEBOOK_IP smoothness.top_25
Lets break down this command a bit:
tools/perf is where we keep our rendering benchmarks. It contains benchmarks, which are written in Python.
run_benchmark is the script we use to run a specific benchmark. Benchmarks are a combination of a "measurement" (e.g. numbers to compute) and a list of pages to compute those numbers for.
--browser=canary tells the script to use Chrome Canary, if it is installed on the system. If you don't have canary [e.g. you're on Linux] it'll fail and tell you to give it another browser.
--browser=list - for all browsers that the script thinks it can use. Pass --browser=list -vvv if you're not seeing a browser you expect to see.
--browser=system - the stable chrome install on your system
release - chromium from out/Debug or out/Release, if it was found
--browser=content-shell-debug - a content shell build found in out/Debug
--browser=android-chrome - chrome detected on an attached android device via adb
--browser=cros-chrome --remote=$CHROMEBOOK_IP - chrome running on your chromebook
--browser=exact --browser-executable=<path to build> - your tests will work with any chrome build >= M18!
smoothness.top_25 is one of our canned benchmarks to run. If you type
./run_benchmark, you'll see a list of other benchmarks that we support. There are a lot, and you can add more. :)
tools/perf/page_sets/top_25.py is a list of 25 pages that get a tremendous amount of volume on desktop, and hence are important for us to not regress. This pageset is used by a variety of benchmarks, from smoothness to thread times. There are other sets of pages too, for example
tough_scrolling_cases. Some have hundreds or thousands of sites. Some have only a few.
- If you want to have a benchmark that computes smoothness, but for a different pageset, then just copy an existing benchmark's .py file and make a new one from it that points at your favorite pageset.
When you run a benchmark, you'll get some output that looks like this:
*RESULT frame_times: frame_times= [16.900000,17.280000,17.340000,16.810000,17.400000,17.500000,17.020000,16.820000,17.070000,17.190000,17.560000,17.670000,17.450000,17.570000,17.270000,17.410000,17.590000,17.530000,17.240000,17.550000,17.190000,17.140000,17.090000,17.500000,17.540000,17.000000,17.300000,17.600000,17.430000,17.070000,17.070000,17.760000,17.090000,16.950000,17.020000,17.040000,16.780000,17.060000,17.700000,17.850000,17.230000,17.090000,17.110000,17.110000,17.610000,17.200000,16.990000,17.180000,17.140000,17.130000,17.430000,17.080000,17.100000,17.100000,17.970000,17.150000,17.600000,17.400000,17.140000,16.920000,17.790000,16.780000,17.440000,16.860000,17.720000,17.700000,17.610000,16.940000,17.200000,16.980000,17.260000,17.310000,17.380000,16.960000,17.000000,17.500000,17.240000,17.170000] ms
Avg frame_times: 17.267564ms
Sd frame_times: 0.279332ms
*RESULT jank: jank= 19.4673
*RESULT mean_frame_time: mean_frame_time= 17.268 ms
*RESULT mostly_smooth: mostly_smooth= 1.0
RESULT telemetry_page_measurement_results: num_failed= 0 count
RESULT telemetry_page_measurement_results: num_errored= 0 count
These are some key statistics for that page as it scrolled, in the default mode for that platform. But, lets say you wanted to run chrome with some extra commandline arguments. For this,
--extra-browser-args is your friend:
Fun! Remember, unless you pass
tools/perf/run_benchmark --browser=canary smoothness.top_25 tools/perf/page_sets/top_25.json --extra-browser-args="--force-llamas-to-have-parachutes --use-rainbows-instad-of-omnibox"
--disable-gpu-vsync, scrolling goes only as fast as your screen. So, for screen with 60 Hz refresh, 16.6 is usually a good thing.
These are the most important smoothness metrics:
- mean_frame_time: arithmetic mean of frame times
- jank: absolute discrepancy of frame time stamps, where discrepancy is a measure of irregularity. It quantifies the worst jank. For a single pause, discrepancy corresponds to the length of this pause in milliseconds. Consecutive pauses increase the discrepancy. The metric is important because even if the mean and 95th percentile are good, one long pause in the middle of an interaction is still bad.
- mostly_smooth: were 95 percent of the frames hitting 60 fps; boolean value (1/0).
- frame_times: list of raw frame times, helpful to understand the above 3 metrics
- mean_pixels_approximated: percentage of pixels that was approximated (checkerboarding, low-resolution tiles, etc.).
Throughout the metrics, you will see the words paint, record, and raster. These have very precise meanings:
- paint: Time dumping WebKit's rendering structures into the compositor's rendering structures in software mode and regular compositing modes. This is the time spent to walk the webkit tree AND software-rasterize its 2D ops AND any time required to do image decodes
- record: (impl-side painting mode only). This is the time to JUST walk the WebKit tree and dump it into an SkPicture.
- raster: (impl-side painting mode only). This is the time to rasterize SkPictures to tiles. If we had a decode cache miss, will include time servicing the image cache miss.
In some sense impl-side painting splits paint into raster + record.
rasterize_and_record measurement calculates the time spent in raster and record. It automatically enables impl-side painting and only works on platforms that support this feature.
How it works
Page scrolling is done by telemetry's "scroll" action,
tools/telemetry/telemetry/internal/actions/scroll.py. On chrome, it boots the browser with
--enable-gpu-benchmarking-extension, which exposes a
Other synthetic gestures, such as pinch-to-zoom or swipe, are available. See the Synthetic Gestures in Chrome document for a full list of gestures and how they're implemented in Chrome.
The smoothness measurement
tools/perf/measurements/smoothness.py captures a trace from this interaction, and extracts the time stamps of frames that were generated. Specifically, the
BenchmarkInstrumentation::MainThreadRenderingStats events are analyzed. These events are issued from
cc/debug/benchmark_instrumentation.cc based on data collected through
cc/debug/rendering_stats_instrumentation.h. The smoothness metric
tools/perf/metrics/smoothness.py is then used to calculate the metrics described above (mean_fram_time, jank, etc.).
Telemetry provides a way to separate out the measurement process from the interaction process from the actual pages being tested. We then maintain a number of important lists of web pages (page sets), some synthetic some real, in
tools/perf/page_sets, grouped by their kind of importance.
key_mobile_sites are likely of particular interest to users.
Telemetry provides a mechanism to very reliably record a web page and then replay it many times in that exact recorded state. We (Chrome team) cannot make our recordings public since the assets the recording are the property of the site owners. However, we have exposed a utility that anyone can use to make their own recordings:
tools/perf/record_wpr --browser=system tools/perf/page_sets/top_25.json
This will place a file called
tools/perf/page_sets/data that is an archive of the data required to replay those pages back over-and-over again without deviation.
Adding credentials to test live sites that require a logged in user
As part of GPU testing, we often want to measure the performance of a site like Gmail, or Facebook, that sit behind a login. We do not give out logins for these, but if you have your own, you can put a
~/.telemetry-credentials in the style of
tools/telemetry/examples/credentials_example.json with the right logins and telemetry will automatically then login to gmail or facebook for you. Patches are welcome to add support for other sites as well.