It is well known that GraalVM native-image can give you fast startup. However, the peak execution speed may lack behind JVM with JIT compiler (see great explanation of such behavior). In short, the missing execution profiles are the problem. However, native-image enterprise is capable to collect and use such profiles for its ahead-of-time compilation. This demo shows the speedup obtained by collecting execution profiles and re-running the native-image compilation with the gathered profiling data.

Get GraalVM Enterprise Edition version 19.0.2 by downloading it from the Oracle Technology Network. Unpack it.

Download Oracle GraalVM Enterprise Edition Native Image JAR file from the same OTN page. Install it using the gu tool:

$ /graalvm-ee-19.0.2/bin/gu install --file native-image-installable-svm-svmee-*-19.0.2.jar

Verify that PGO support is available:

$ /graalvm-ee-19.0.2/bin/native-image --help | grep pgo
    --pgo                 a comma-separated list of files from which to read the data
    --pgo-instrument      instrument AOT compiled code to collect data for profile-guided

With the GraalVM EE 19.0.2 properly installed we can proceed with initial measuring:

$ JAVA_HOME=/graalvm-ee-19.0.2/ mvn clean install
$ ./target/geom 15000 now 30 square rectangle
sum: 4394.341496946556
last round 55 ms
$ ./target/geom 15000 now 30 triangle circle
sum: 8776.729175399201
last round 48 ms

That is the initial benchmark we want to beat. The default code is ready for four different objects square, rectangle, triangle and circle and treats all of them with the same probability.

However we can train the code to optimize for certain objects. To do that we have to generate a special version of the binary:

$ JAVA_HOME=/graalvm-ee-19.0.2/ mvn clean install -PProfilesCollect
$ ./target/geom 15000 now 30 square rectangle
...
sum: 4384.131082976915
last round 108 ms
    $ ls *iprof
default.iprof

Now the same algorithm runs slower, however that is because we are collecting the profiling data. Once the program is finished, an .iprof file is generated.

Now we can recompile once again and see the speedup. Do install again. This time it finds the geom.iprof file and optimizes the binary:

$ JAVA_HOME=/graalvm-ee-19.0.2/ mvn install -PProfilesUse
$ ./target/geom 15000 now 30 square rectangle
...
sum: 4401.419731937973
last round 36 ms

That means 35% speedup!

However, the speed is only improved when the trained for geometric shapes are being processed. Should the other ones appear too, the execution slows down. At the end, it can be even less performant than the non-optimized version:

$ ./target/geom 15000 now 30 triangle circle
sum: 8888.679818384859
last round 74 ms

If something like this happens, it is time to re-profile and re-deploy new version.

It is well known that native-image can give you fast startup. However with the help of the --pgo-instrument and --pgo options, it can also give you execution speed optimal for your workloads.