A tool to turn the output of -XX: TraceBytecodes (a JDK debug-only feature to print every bytecode executed by the interpreter) into a simple stack format, and a script to turn that into flame graphs showing what happened during program execution, bytecode-by-bytecode:
This is a diagnostic tool I've developed and used to visualize what is going on at startup at a high granularity. Since it relies on a very verbose and slow debug option it's not by any means a good estimator of how long time things actually take.
As such, this is more a diagnostic tool than a profiler, but a tool that has served us well to disentangle various startup dependencies and inefficiencies in the JDK itself - and been a valuable aid in diagnosing a number of startup regressions.
Most profiling tools either have a blind spot when it comes to capturing what happens during the very early initialization of a JVM, e.g., agent based instrumentation, JFR etc, or they are native profiling tools that would attribute everything to "something in the interpreter".
- git and perl required to get FlameGraph working
- a fastdebug (recommended) or slowdebug build of OpenJDK (you may have to build one yourself)
If you are on Linux, you can download a fastdebug binary from https://www.chriswhocodes.com/ and install it with Jabba:
jabba install fastdebug@11=tgz https://www.chriswhocodes.com/downloads/jdk11-chriswhocodes-nightly-linux-x86_64-fastdebug.tgz
jabba use fastdebug@11
Jabba also supports XZ format, so you can download nightly and 15 builds:
jabba install shipilev-fastdebug@nightly=tgx https://builds.shipilev.net/openjdk-jdk/openjdk-jdk-latest-linux-x86_64-fastdebug.tar.xz
jabba install shipilev-fastdebug@15=tgx https://builds.shipilev.net/openjdk-jdk15/openjdk-jdk15-latest-linux-x86_64-fastdebug.tar.xz
- Build:
./gradlew build
- With a fastdebug build of java, run any (small) program with
-XX: TraceBytecodes
, e.g.:java -XX: TraceBytecodes HelloWorld | ./bytestacks helloworld
- This generates
helloworld.stacks
andhelloworld.svg
only.
- This generates
- open
helloworld.svg
in your favorite SVG viewer. A modern web browser should suffice.
By not retaining the trace output we'll run a lot faster and we avoid wasting disk space, but if you need the tracing output it's easy to split the steps apart:
- Run
java -XX:TraceBytecodes > helloworld
(this dumps the raw tracing output tohelloworld
- Run
./bytestacks helloworld
(this generateshelloworld.stacks
and usesFlameGraph
to producehelloworld.svg
)
- To get the full picture of what's going on, especially on a longer running program, it might be interesting to run your program with
-Xint
, since as things start to get compiled they will disappear from view... - For performance reasons there is a granularity option which allows setting a threshold for how small methods are output in the stack output, defaulting to 25. By setting this to a lower value, the generated flame graph will become more detailed, but may also become very heavy to render