This repo combines an implementation of JEP candidate with an async-profiler fork which uses the API, giving an example on how to easily migrate an application which uses the old API.

Either build the JDK in the folder jdk as you would usually do (it's a JDK 20, release builds are recommended) and build the async-profiler in the folder async-profiler via make or run ./build.sh. Be sure to install the required dependencies (you will probably see related error messages if you don't) and use a current JDK (19 or 20).

It is based on OpenJDK head but the changes should be easy to backport to previous versions.

./run.sh AGENT_ARGS JAVA_ARGS... which uses the built JDK and async-profiler.

For example, to run a dacapo benchmark, e.g jython, and generate a flame graph run

test -e dacapo.jar || wget https://downloads.sourceforge.net/project/dacapobench/9.12-bach-MR1/dacapo-9.12-MR1-bach.jar -O dacapo.jar
./run.sh flat=10,traces=1,interval=500us,event=cpu,flamegraph,file=flame.html -XX:+UnlockDiagnosticVMOptions -XX:+DebugNonSafepoints -jar dacapo.jar jython
open flame.html

With an interval of 500us (0.5ms), more information on the arguments in the async-profiler. Use another benchmark like tomcat instead of jython, if the flame graph misses the bottom frames.

This resulted in a flame graph like (click on the image to get to the HTML flame graph):

Orange frames are related to C/C++ internal JVM methods, red frames are related to other C/C++ code, darker green frames to interpreted methods, lighter green frames to compiled methods and blue frames to inlined compiled methods.

The usage of the new AsyncGetStackTrace gives us the following additions to a normal async-profiler flame graph: Information on the compilation stage (C1 vs C2 compiler), inlining information for non-top frames, information on C frames starting with _pthread_start up to the first Java frame and C frames between Java frames. This information was previously unobtainable by async-profiler (or any other profiler using just JFR or AsyncGetCallTrace).

The same flame graph using the old AsyncGetCallTrace can be generated using the following:

test -e dacapo.jar || wget https://downloads.sourceforge.net/project/dacapobench/9.12-bach-MR1/dacapo-9.12-MR1-bach.jar -O dacapo.jar
test -e ap-loader.jar || wget https://github.com/jvm-profiling-tools/ap-loader/releases/latest/download/ap-loader-all.jar -O ap-loader.jar
java -javaagent:./ap-loader.jar=start,flat=10,traces=1,interval=500us,event=cpu,flamegraph,file=flame_old.html -jar dacapo.jar jython
open flame_old.html

This resulted in a flame graph like (click on the image to get to the HTML flame graph):

Showing the stark difference in the amount of available information.