Java FFI benchmarks

Benchmarks for Java JNI vs Project Panama on Linux (Fedora 35) with JDK (openjdk 17-panama 2021-09-14)

Setup instructions for benchmark:

Clone the project

git clone

Setup Java 17 panama build using SDK man

sdk install java 17.ea.3.pma-open
sdk use java 17.ea.3.pma-open

Create Java bindings for `unistd.h`

Linux

jextract --source -d generated/src/main/java -t org.unix -I /usr/include /usr/include/unistd.h

MacOS

export C_INCLUDE=/Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include
jextract --source -d generated/src/main/java -t org.unix -I $C_INCLUDE $C_INCLUDE/unistd.h

Build and run using Maven

When running the benchmark it currently only works for Linux.

mvn clean verify 
java -jar target/benchmarks.jar

Typical Results

These results vary because they were ran on a developer machine (macbook pro) with other services running. Also, according to the Java Microbenchmark Harness docs, to avoid blackholes (methods that return void). e.g. If you call getPid() and return void the JVM will optimize by removing dead code. To ensure the code isn't removed the method returns a primitive (int). Below smaller numbers the better.

Benchmark                    Mode  Cnt  Score   Error  Units
FFIBenchmark.JNI             avgt   40  9.698 ± 0.532  ns/op
FFIBenchmark.panamaDowncall  avgt   40  8.431 ± 0.096  ns/op
FFIBenchmark.panamaJExtract  avgt   40  8.488 ± 0.099  ns/op

40 count

Benchmark                    Mode  Cnt  Score   Error  Units
FFIBenchmark.JNI             avgt   40  49.182 ± 1.079  ns/op
FFIBenchmark.panamaDowncall  avgt   40  50.746 ± 0.702  ns/op
FFIBenchmark.panamaJExtract  avgt   40  48.838 ± 1.461  ns/op