These are instructions for collecting a CPU profile of chromium. All of the profiling methods described here produce output that can be view using the pprof tool. pprof is highly customizable; here's a screenshot of some example pprof output:
This doc is intended to be an authoritative one-stop resource for profiling chromium. At the time of writing, there are a number of existing docs with profiling instructions, in varying states of obsolescence:
CPU profiling is not to be confused with tracing or task profiling:
Profiling should preferably be done on an official build. Make sure the following appears in your args.gn file:
is_official_build = true
First, make sure you have the linux-perf package installed:
$ sudo apt-get install linux-perf
After starting up the browser and loading the page you want to profile, press ‘Shift-Escape’ to bring up the task manager, and get the Process ID of the process you want to profile.
Run the perf tool like this:
$ perf record -g -p <Process ID> -o <output file>
-F argument, e.g., -F 1000.-e cpu-clock.To stop profiling, press Control-c in the terminal window where perf is running. Run pprof to view the results, providing the path to the browser executable; e.g.:
$ pprof -web src/out/Release/chrome <perf output file>
pprof is packed with useful features for visualizing profiling data. Try pprof --help for more info.gcert first will make pprof run faster, and eliminate some useless spew to the terminal.If you want to profile all renderer processes use the custom --renderer-cmd-prefix profiling script:
$ src/out/Release/chrome --renderer-cmd-prefix=“tools/profiling/linux-perf-renderer-cmd.sh”
If you want to limit the profile to a single thread, run:
$ ps -T -p <Process ID>
From the output, find the Thread ID (column header “SPID”) of the thread you want. Now run perf:
$ perf record -g -t <Thread ID> -o <output file>
Use the same pprof command as above to view the single-thread results.
You can generate a highly-focused profile for any period that can be defined in javascript using the chrome.gpuBenchmarking javascript interface. First, adding the following command-line flags when you start chrome:
$ chrome --enable-gpu-benchmarking --no-sandbox [...]
Open devtools, and in the console, use chrome.gpuBenchmarking.startProfiling and chrome.gpuBenchmarking.stopProfiling to define a profiling period. e.g.:
> chrome.gpuBenchmarking.startProfiling('perf.data'); doSomething(); chrome.gpuBenchmarking.stopProfiling()
chrome.gpuBenchmarking has a number of useful methods for simulating user-gesture-initiated actions; for example, to profile scrolling:
> chrome.gpuBenchmarking.startProfiling('perf.data'); chrome.gpuBenchmarking.smoothScrollByXY(0, 1000, () => { chrome.gpuBenchmarking.stopProfiling() });
This section contains instructions on how to do profiling using the callgrind/cachegrind tools provided by valgrind. This is not a sampling profiler, but a profiler based on running on a simulated CPU. The instructions are Linux-centered, but might work on other platforms too.
sudo apt-get install valgrind
Run content_shell with callgrind to create a profile. A callgrind.<pid> file will be dumped when exiting the browser or stopped with CTRL-C:
valgrind --tool=callgrind content_shell --single-process --no-sandbox <url>
Alternatively use cachegrind which will give you CPU cycles per code line:
valgrind --tool=cachegrind content_shell --single-process --no-sandbox <url>
Using single-process is for simple profiling of the renderer. It should be possible to run in multi-process and attach to a renderer process.
Warning: this will install a bunch of KDE dependencies.
sudo apt-get install kcachegrind
kcachegrind callgrind.<pid>
Android (Nougat and later) supports profiling using the simpleperf tool.
Follow the instructions for building and installing chromium on android. With chromium running on the device, run the following command to start profiling on the browser process (assuming your build is in src/out/Release):
$ src/out/Release/bin/chrome_public_apk profile Profiler is running; press Enter to stop...
Once you stop the profiler, the profiling data will be copied off the device to the host machine and post-processed so it can be viewed in pprof, as described above.
To profile the renderer process, you must have just one tab open in chromium, and use a command like this:
$ src/out/Release/bin/chrome_public_apk profile --profile-process=renderer
To limit the profile to a single thread, use a command like this:
$ src/out/Release/bin/chrome_public_apk profile --profile-process=renderer --profile-thread=main
The --profile-process and --profile-thread arguments support most of the common process names (‘browser’, ‘gpu’, ‘renderer’) and thread names (‘main’, ‘io’, ‘compositor’, etc.). However, if you need finer control of the process and/or thread to profile, you can specify an explicit Process ID or Thread ID. Check out the usage message for more info:
$ src/out/Release/bin/chrome_public_apk help profile
By default, simpleperf will collect CPU cycles. To collect other events, use a command like this:
$ src/out/Release/bin/chrome_public_apk profile --profile-process=renderer --profile-events cpu-cycles,branch-misses,branch-instructions,cache-references,cache-misses,stalled-cycles-frontend,stalled-cycles-backend
Follow the simple chrome instructions, to build and deploy chrome to your chromeos device. These instructions will set up a build directory for you, so be sure to gn args out_${SDK_BOARD}/Release to edit them and add the gn args listed above.
The easiest way to get a profile is to ssh to your device, which here will be referred to as chromeos-box, but replace that with whatever ip or hostname your device is. ssh to your device, create a folder in /tmp (which usually has more space than /) and record performance for the entire device. When you're done, use scp to copy the perf.data back to your desk and use pprof as per normal on that perf.data file.
Here's an example:
$ ssh root@chromeos-box
localhost ~ # export CPUPROFILE_FREQUENCY=3000
localhost ~ # mkdir -p /tmp/perf
localhost ~ # cd /tmp/perf
localhost /tmp/perf # perf record -g -a -e cycles
^C
[ perf record: Woken up 402 times to write data ]
[ perf record: Captured and wrote 100.797 MB perf.data (489478 samples) ]
localhost /tmp/perf # exit
$ scp root@chromeos-box:/tmp/perf/perf.data .
$ pprof -web out_${SDK_BOARD}/Release/chrome perf.data
Note: this will complain about missing chromeos symbols. Even pointing PPROF_BINARY_PATH at the expanded debug-board.tgz file that came along with the chromeos image does not seem to work. If you can make this work, please update this doc!
The perf benchmark runner can generate a CPU profile over the course of running a perf test. Currently, this is supported only on Linux and Android. To get info about the relevant options, run:
$ src/tools/perf/run_benchmark help run
... and look for the --interval-profiling-* options. For example, to generate a profile of the main thread of the renderer process during the “page interactions” phase of a perf benchmark, you might run:
$ src/tools/perf/run_benchmark run <benchmark name> --interval-profiling-target=renderer:main --interval-profiling-period=interactions --interval-profiling-frequency=2000
The profiling data will be written into the artifacts/ sub-directory of your perf benchmark output directory (default is src/tools/perf), to files with the naming pattern *.profile.pb. You can use pprof to view the results, as described above.
If you use pprof -proto chrome-profile-renderer-12345 to turn your perf data into a proto file, you can then use that resulting file with internal tools. See http://go/cprof/user#fs-profiles for instructions on how to go about this.
Many of the profiling tools expect you to provide the PID of the process to profile. If the tool used does not support finding the application by name or you would like to run the command for many processes it can be useful to use pgrep to find the PIDs.
Find the PID for Chromium (browser process):
$ pgrep -X Chromium
Find the PID for all child processes of Chromium:
$ pgrep -P $CHROMIUM_PID
Combine commands to run tool for Chromium and all its children:
$ cat <(pgrep -x Chromium) <(pgrep -P $(pgrep -x Chromium)) | xargs $MY_TOOL --pid
Profiling should always be done on a build that represents the performance of official builds as much as possible. is_official_build enables some additional optimizations like PGO.
is_debug = false is_component_build = false is_official_build = true # Most profiling techniques on macOS will work with minimal symbols for local builds. # You should try and use minimal symbols when starting out because most tools will take # an incredibly long time to process the symbols and in some cases will freeze the application # while doing so. symbol_level sets the level for all parts of Chromium. The # blink and v8 settings allow overriding this to set higher or lower levels # for those components. blink_symbol_level = 0 v8_symbol_level = 0 symbol_level = 0
Once collected the traces produced by any tool in this section can be converted to pprof using InstrumentsToPprof.
Sample stacks of $pid for 10 seconds grabbing a stack every 1ms. [-maydie] to still have stacks if process exits.
$ sample $pid 10 1 -mayDie -f ./output.txt
To get a trace use either the GUI in the “Time Profiler” mode or this command:
$ xcrun -r xctrace record --template 'Time Profiler' --all-processes --time-limit 30s --output 'profile.trace'
By default dtrace does not work well with SIP. Disabling SIP as a whole is not recommended and instead should be done only for DTrace using these steps:
csrutil enable --without dtrace --without debugTo get sampled cpu stacks
$ dtrace -p $PID -o $OUTPUT_FILE -n "profile-1001/pid == $PID/ {{ @[ustack()] = count(); }}"
To get stacks that caused wake-ups
$ dtrace -p $PID -o $OUTPUT_FILE -n "mach_kernel::wakeup/pid == $PID/ {{ @[ustack()] = count(); }}"