Saint's Log

Experimenting with Async Profiler

I have been studying the performance of a simple Java application (for integer factorization) using async-profiler. The application’s source code is on GitHub.

async-profiler is a low overhead sampling profiler for Java that does not suffer from Safepoint bias problem.

async-profiler repo

There is also a 3-part talk about async-profiler demo-ing how it works and how to use it.

I downloaded the Linux x64 async-profiler build with these commands…

mkdir -p ~/java/binaries/async-profiler
cd ~/java/binaries/async-profiler

curl -Lo async-profiler-2.9-linux-x64.tar.gz https://github.com/jvm-profiling-tools/async-profiler/releases/download/v2.9/async-profiler-2.9-linux-x64.tar.gz

tar xzf async-profiler-2.9-linux-x64.tar.gz

… then started the application with these:

# macos:
export JAVA_HOME=~/java/binaries/jdk/x64/jdk-17.0.7+7/Contents/Home

# Linux:
export JAVA_HOME=~/java/binaries/jdk/x64/jdk-17.0.7+7/

cd ~/repos/scratchpad/demos/java/FindPrimes/
$JAVA_HOME/bin/java Factorize 91278398257125987

Once the application is running, use the profiler.sh script to attach to the Java process and start profiling it. I was interested in wall clock profiling. This is specified using the -e wall argument (see Part 2: Improving Performance with Async-profiler by Andrei Pangin. – YouTube). The command line below will profile the Java application with a 5ms sampling interval for a duration (-d) of 10 seconds.

# macos:
cd ~/java/binaries/async-profiler-2.9-macos

# Linux:
cd ~/java/binaries/async-profiler-2.9-linux

./profiler.sh -e wall -t -i 5ms -d 10 -f result.html jps

The jps argument above lets the profiler.sh script determine which Java process is running by calling The jps Command (oracle.com). If there are multiple Java processes, then run jps first to determine the process id of the one to be profiled then explicitly pass that pid to profiler.sh e.g.

jps
./profiler.sh -e wall -t -i 5ms -d 10 -f result.html 53361

Async-profiler can also be attached at application startup.

$JAVA_HOME/bin/java -agentlib:~/java/binaries/async-profiler-2.9-macos/build/libasyncProfiler.dylib=start,event=wall,threadsfile=out.html Factorize

Other Event Types

The event type we have used so far is the wall clock event. Other event types include cpu and lock. The latter mode is mentioned at 32:15 in Part 2: Improving Performance with Async-profiler by Andrei Pangin. – YouTube. Here’s a sample command:

~/java/binaries/async-profiler/async-profiler-2.9-linux-x64/profiler.sh -e lock -i 5ms -d 10 -f result.html jps

Output Formats

The profiling data can be written in various formats. Here is an example command line used to explore what the traces output looks like. See scratchpad/factorization-profile.sh · swesonga/scratchpad · GitHub for additional examples of output formats.

cd ~/java/binaries/async-profiler/async-profiler-2.9-linux-x64
./profiler.sh -e wall -i 5ms -d 10 -o flat,traces -f Factorize-flat+traces.html 69490

To convert the async-profiler data from one format to another, use converter.jar from the downloaded tar file:

$JAVA_HOME/bin/java -cp ~/java/binaries/async-profiler/async-profiler-2.9-linux-x64/build/converter.jar FlameGraph rawdata.txt output.html

Other Notes

To find out file types on macos, run file -I rawdata. In my case, I had flamegraph data that was shared as application/gzip (causing unzip to fail with End-of-central-directory signature not found. I needed to use gzip -d rawdata.