Profiling Ray

This document details, for Ray developers, how to use pprof to profile Ray binaries.


These instructions are for Ubuntu only. Attempts to get pprof to correctly symbolize on Mac OS have failed.

sudo apt-get install google-perftools libgoogle-perftools-dev

Changes to compilation and linking

Let’s say we want to profile the plasma_manager. Change the link instruction in src/plasma/CMakeLists.txt from

target_link_libraries(plasma_manager common ${PLASMA_STATIC_LIB} ray_static ${ARROW_STATIC_LIB} -lpthread)

to additionally include -lprofiler:

target_link_libraries(plasma_manager common ${PLASMA_STATIC_LIB} ray_static ${ARROW_STATIC_LIB} -lpthread -lprofiler)

Additionally, add -g -ggdb to CMAKE_C_FLAGS and CMAKE_CXX_FLAGS to enable the debug symbols. (Keeping -O3 seems okay.)


Launching the to-profile binary

In various places, instead of launching the target binary via plasma_manager <args>, it must be launched with

LD_PRELOAD=/usr/lib/ CPUPROFILE=/tmp/pprof.out plasma_manager <args>

In practice, this means modifying python/ray/plasma/ so that the manager is launched with a command that passes a modified_env into Popen.

modified_env = os.environ.copy()
modified_env["LD_PRELOAD"] = "/usr/lib/"
modified_env["CPUPROFILE"] = "/tmp/pprof.out"

process = subprocess.Popen(command,

The file /tmp/pprof.out will be empty until you let the binary run the target workload for a while and then kill it.

Visualizing the CPU profile

The output of pprof can be visualized in many ways. Here we output it as a zoomable .svg image displaying the call graph annotated with hot paths.

# Use the appropriate path.

google-pprof -svg $PLASMA_MANAGER /tmp/pprof.out > /tmp/pprof.svg
# Then open the .svg file with Chrome.

# If you realize the call graph is too large, use -focus=<some function> to zoom
# into subtrees.
google-pprof -focus=epoll_wait -svg $PLASMA_MANAGER /tmp/pprof.out > /tmp/pprof.svg

Here’s a snapshot of an example svg output, taken from the official documentation: