gProfiler

gProfiler combines multiple sampling profilers to produce unified visualization of what your CPU is spending time on, displaying stack traces of your processes across native programs¹ (includes Golang), Java and Python runtimes, and kernel routines.

gProfiler can upload its results to the Granulate Performance Studio, which aggregates the results from different instances over different periods of time and can give you a holistic view of what is happening on your entire cluster. To upload results, you will have to register and generate a token on the website.

gProfiler runs on Linux.

Running

This section describes the possible options to control gProfiler's output, and the various execution modes (as a container, as an executable, etc...)

Output options

gProfiler can produce output in two ways:

• Create an aggregated, collapsed stack samples file (profile_<timestamp>.col) and a flamegraph file (profile_<timestamp>.html). Two symbolic links (last_profile.col and last_flamegraph.html) always point to the last output files.

  Use the --output-dir/-o option to specify the output directory.

  If --rotating-output is given, only the last results are kept (available via last_profle.col and last_flamegraph.html). This can be used to avoid increasing gProfiler's disk usage over time. Useful in conjunction with --upload-results (explained ahead) - historical results are available in the Granulate Performance Studio, and the very latest results are available locally.

  --no-flamegraph can be given to avoid generation of the profile_<timestamp>.html file - only the collapsed stack samples file will be created.

• Send the results to the Granulate Performance Studio for viewing online with filtering, insights, and more.

  Use the --upload-results/-u flag. Pass the --token option to specify the token provided by Granulate Performance Studio, and the --service-name option to specify an identifier for the collected profiles, as will be viewed in the Granulate Performance Studio. Profiles sent from numerous gProfilers using the same service name will be aggregated together.

Note: both flags can be used simultaneously, in which case gProfiler will create the local files and upload the results.

Profiling options

• --profiling-frequency: The sampling frequency of the profiling, in hertz.
• --profiling-duration: The duration of the each profiling session, in seconds.
• --profiling-interval: The interval between each profiling session, in seconds.

The default profiling frequency is 11 hertz. Using higher frequency will lead to more accurate results, but will create greater overhead on the profiled system & programs.

The default duration is 60 seconds, and the default interval matches it. So gProfiler runs the profiling sessions back-to-back - the next session starts as soon as the previous session is done.

Java profiling options

• --no-java: Disable profilers for Java.

Python profiling options

• --no-python: Alias of --python-mode none.
• --python-mode: Controls which profiler is used for Python.
  • auto - (default) try with PyPerf (eBPF), fall back to py-spy.
  • pyperf - Use PyPerf with no py-spy fallback.
  • pyspy - Use py-spy.
  • none - Disable profilers for Python.

Profiling using eBPF incurs lower overhead & provides kernel stacks. This (currently) requires kernel headers to be installed.

PHP profiling options

• --no-php: Disable profilers for PHP.
• --php-proc-filter: Process filter (pgrep) to select PHP processes for profiling (this is phpspy's -P option)

Ruby profiling options

• --no-ruby: Disable profilers for Ruby.

System profiling options

• --perf-mode: Controls the global perf strategy. Must be one of the following options:
  • fp - Use Frame Pointers for the call graph
  • dwarf - Use DWARF for the call graph (adds the --call-graph dwarf argument to the perf command)
  • smart - Run both fp and dwarf, then choose the result with the highest average of stack frames count, per process.
  • none - Avoids running perf at all. See perf-less mode.

Continuous mode

gProfiler can be run in a continuous mode, profiling periodically, using the --continuous/-c flag. Note that when using --continuous with --output-dir, a new file will be created during each sampling interval. Aggregations are only available when uploading to the Granulate Performance Studio.

Running as a Docker container

Run the following to have gProfiler running continuously, uploading to Granulate Performance Studio:

docker pull granulate/gprofiler:latest
docker run --name gprofiler -d --restart=always \
    --pid=host --userns=host --privileged \
    -v /lib/modules:/lib/modules:ro -v /usr/src:/usr/src:ro \
    -v /var/run/docker.sock:/var/run/docker.sock \
	granulate/gprofiler:latest -cu --token <token> --service-name <service> [options]

For profiling with eBPF, kernel headers must be accessible from within the container at /lib/modules/$(uname -r)/build. On Ubuntu, this directory is a symlink pointing to /usr/src. The command above mounts both of these directories.

Running as an executable

Run the following to have gprofiler running continuously, uploading to Granulate Performance Studio:

wget https://github.com/Granulate/gprofiler/releases/latest/download/gprofiler
sudo chmod +x gprofiler
sudo ./gprofiler -cu --token <token> --service-name <service> [options]

gProfiler unpacks executables to /tmp by default; if your /tmp is marked with noexec, you can add TMPDIR=/proc/self/cwd to have everything unpacked in your current working directory.

sudo TMPDIR=/proc/self/cwd ./gprofiler -cu --token <token> --service-name <service> [options]

Executable known issues

The following platforms are currently not supported with the gProfiler executable:

• Alpine

Remark: container-based execution works and can be used in those cases.

Running as a Kubernetes DaemonSet

See gprofiler.yaml for a basic template of a DaemonSet running gProfiler. Make sure to insert the GPROFILER_TOKEN and GPROFILER_SERVICE variables in the appropriate location!

Running from source

gProfiler requires Python 3.6+ to run.

pip3 install -r requirements.txt
./scripts/copy_resources_from_image.sh

Then, run the following as root:

python3 -m gprofiler [options]

Theory of operation

Each profiling interval, gProfiler invokes perf in system wide mode, collecting profiling data for all running processes. Alongside perf, gProfiler invokes runtime-specific profilers for processes based on these environments:

• Java runtimes (version 7+) based on the HotSpot JVM, including the Oracle JDK and other builds of OpenJDK like AdoptOpenJDK and Azul Zulu.
  • Uses async-profiler.
• The CPython interpreter, versions 2.7 and 3.5-3.9.
  • eBPF profiling (based on PyPerf) requires Linux 4.14 or higher; see Python profiling options for more info.
  • If eBPF is not available for whatever reason, py-spy is used.
• PHP (Zend Engine), versions 7.0-8.0.
  • Uses Granulate's fork of the phpspy project.
• Ruby versions (versions 1.9.1 to 3.0.1)
  • Uses Granulate's fork of the rbspy profiler.

The runtime-specific profilers produce stack traces that include runtime information (i.e, stacks of Java/Python functions), unlike perf which produces native stacks of the JVM / CPython interpreter. The runtime stacks are then merged into the data collected by perf, substituting the native stacks perf has collected for those processes.

perf-less mode

It is possible to run gProfiler without using perf - this is useful where perf can't be used, for whatever reason (e.g permissions). This mode is enabled by --perf-mode none.

In this mode, gProfiler uses runtime-specific profilers only, and their results are concatenated (instead of scaled into the results collected by perf). This means that, although the results from different profilers are viewed on the same graph, they are not necessarily of the same scale: so you can compare the samples count of Java to Java, but not Java to Python.

Contribute

We welcome all feedback and suggestion through Github Issues:

• Submit bugs and feature requests
• Upvote popular feature requests

Releasing a new version

1. Update __version__ in __init__.py.
2. Create a tag with the same version (after merging the __version__ update) and push it.

We recommend going through our contribution guide for more details.

Credits

• async-profiler by Andrei Pangin. See our fork.
• py-spy by Ben Frederickson. See our fork.
• bcc (for PyPerf) by the IO Visor project. See our fork.
• phpspy by Adam Saponara. See our fork.
• rbspy by the rbspy project. See our fork

Footnotes

1: To profile native programs that were compiled without frame pointers, make sure you use the --perf-mode smart (which is the default). Read more about it in the Profiling options section↩