Java Flame Graphs
Java flame graphs
A few years ago, Brendan Gregg created the flame graph visualization. He describes it as a visualization of profiled software, allowing the most frequent code-paths to be identified quickly and accurately. More pragmatically, it is a great way to identify the hot spots of an application and to understand its runtime behavior.
The FlameGraph project provides a flamegraph.pl script to create SVG files from a pivot format called folded stacks, and several scripts to convert the proprietary output of various profilers (perf, DTrace etc.) into folded stacks.
This project aims to provide conversion scripts for several Java profilers. These scripts are not pushed in the official project mostly because they are written in Python and not in Perl.
Supported profilers are:
The easiest way to install hprof2flamgegraph is to use the Pypi package:
pip install [--user] hprof2flamegraph
It installs a stackcollapse-hprof and stackcollapse-hpl scripts into the bin directory of your environment. Make sure this directory is in your PATH. The original flamegraph.pl script from Brendan is also installed (CDDL licensed).
You can also download the script from github or clone the repository. The script is standalone with Python >= 2.7 and only requires the argparse module with Python 2.6.
Why should I use flame graphs with a Java profiler ?
The Java ecosystem is full of great performance analysis tools and profilers. They are often full featured (CPU, memory, threads, GC, monitors, etc.) and well suited for complex environments or analyses.
However, quite often we only need to get the big picture of our application. Flame graphs really shine in this area. It is the easiest and fastest way to visualize your application and understand its performance profile. It is a complement to more heavy-weight analysis environments.
The official flame graph visualization page describe it in-depth and examine several use cases.
Run the application with HPROF enabled. It must be configured to do CPU sampling and not CPU tracing. You can configure the sampling interval, the maximum stack depth, and whether if line numbers and thread information are printed.
I recommend to always set these last two to y since they can be discarded at a latter step if needed. You never collect too much information.
Convert the output.hprof file into folded stacks using the stackcollapse-hprof script
stackcollapse-hprof output.hprof > output-folded.txt
Create the final SVG graph. You can either use the flamegraph.pl script shipped with this module or the one from the official FlameGraph project. They are the same.
flamegraph.pl output-folded.txt > output.svg
A few tips about HPROF follows:
- HPROF is not hot-pluggable. It means that it must be activated when the JVM starts and that the application will be profiled from the begging to the end. However, playing with SIGQUITs allow to profile a specific time frame. Its lame but it works.
- It is also usually a good practice to avoid round sampling intervals like 10ms to avoid a possible bias due to the synchronization of several periodic processes (like a process scheduler or a timer).
- As shown in Evaluating the Accuracy of Java Profilers HPROF is flawed, like many other Java profiler, and can produce non-actionable profiles. It worth reading the paper to make sure you understand the limitation of the tool. After that, HPROF is usually good enough to identify the hot spots.
Run the application with honest-profiler enabled (and remember it is not production ready yet).
It will create a log.hpl. Convert it into folded stacks using the stackcollapse-hpl script
stackcollapse-hpl log.hpl > output-folded.txt
Create the final SVG graph
flamegraph.pl output-folded.txt > output.svg
Specific use cases
Want to profile an Hadoop job?
It is quite easy to do. You only have to set the following Hadoop variables:
To enable HPROF programmatically from a Java job:
Configuration conf = getConf(); conf.setBoolean("mapred.task.profile", true); conf.set("mapred.task.profile.params", "-agentlib:hprof=cpu=samples,depth=100,interval=7,lineno=y,thread=y,file=%s"); conf.set("mapred.task.profile.maps", "0"); conf.set("mapred.task.profile.reduces", "0");
To do it from the command line:
hadoop jar my.jar \ -Dmapred.task.profile=true \ -Dmapred.task.profile.params="-agentlib:hprof=cpu=samples,depth=100,interval=7,lineno=y,thread=y,file=%s" \ -Dmapred.task.profile.maps=0 \ -Dmapred.task.profile.reduces=0