Optimize a Native Executable with Profile-Guided Optimizations

GraalVM Native Image offers quick startup and less memory consumption for a Java application, running as a native executable, by default. You can optimize this native executable even more for additional performance gain and higher throughput by applying Profile-Guided Optimizations (PGO).

With PGO you can collect the profiling data in advance and then feed it to the native-image tool, which will use this information to optimize the performance of the resulting binary.

Note: PGO is not available in GraalVM Community Edition.

This guide shows how to apply PGO and transform your Java application into an optimized native executable.

Run a Demo

For the demo part, you will run a Java application performing queries implemented with the Java Streams API. A user is expected to provide two integer arguments: the number of iterations and the length of the data array. The application creates the data set with a deterministic random seed and iterates 10 times. The time taken for each iteration and its checksum is printed to the console.

Below is the stream expression to optimize:

Arrays.stream(persons)
   .filter(p -> p.getEmployment() == Employment.EMPLOYED)
   .filter(p -> p.getSalary() > 100_000)
   .mapToInt(Person::getAge)
   .filter(age -> age > 40)
   .average()
   .getAsDouble();

Follow these steps to build an optimized native executable using PGO.

Note: Make sure you have installed a GraalVM JDK. The easiest way to get started is with SDKMAN!. For other installation options, visit the Downloads section.

  1. Save the following code to the file named Streams.java:

    import java.util.Arrays;
import java.util.Random;

public class Streams {

  static final double EMPLOYMENT_RATIO = 0.5;
  static final int MAX_AGE = 100;
  static final int MAX_SALARY = 200_000;

  public static void main(String[] args) {

    int iterations;
    int dataLength;
    try {
      iterations = Integer.valueOf(args[0]);
      dataLength = Integer.valueOf(args[1]);
    } catch (Throwable ex) {
      System.out.println("Expected 2 integer arguments: number of iterations, length of data array");
      return;
    }

    Random random = new Random(42);
    Person[] persons = new Person[dataLength];
    for (int i = 0; i < dataLength; i++) {
      persons[i] = new Person(
          random.nextDouble() >= EMPLOYMENT_RATIO ? Employment.EMPLOYED : Employment.UNEMPLOYED,
          random.nextInt(MAX_SALARY),
          random.nextInt(MAX_AGE));
    }

    long totalTime = 0;
    for (int i = 1; i <= 20; i++) {
      long startTime = System.currentTimeMillis();

      long checksum = benchmark(iterations, persons);

      long iterationTime = System.currentTimeMillis() - startTime;
      totalTime += iterationTime;
      System.out.println("Iteration " + i + " finished in " + iterationTime + " milliseconds with checksum " + Long.toHexString(checksum));
    }
    System.out.println("TOTAL time: " + totalTime);
  }

  static long benchmark(int iterations, Person[] persons) {
    long checksum = 1;
    for (int i = 0; i < iterations; ++i) {
      double result = getValue(persons);

      checksum = checksum * 31 + (long) result;
    }
    return checksum;
  }

  public static double getValue(Person[] persons) {
    return Arrays.stream(persons)
        .filter(p -> p.getEmployment() == Employment.EMPLOYED)
        .filter(p -> p.getSalary() > 100_000)
        .mapToInt(Person::getAge)
        .filter(age -> age >= 40).average()
        .getAsDouble();
  }
}

enum Employment {
  EMPLOYED, UNEMPLOYED
}

class Person {
  private final Employment employment;
  private final int age;
  private final int salary;

  public Person(Employment employment, int height, int age) {
    this.employment = employment;
    this.salary = height;
    this.age = age;
  }

  public int getSalary() {
    return salary;
  }

  public int getAge() {
    return age;
  }

  public Employment getEmployment() {
    return employment;
  }
}
  2. Compile the application: 
    $JAVA_HOME/bin/javac Streams.java

    (Optional) Run the demo application, providing some arguments to observe performance.

    $JAVA_HOME/bin/java Streams 100000 200
  3. Build a native executable from the class file, and run it to compare the performance: 
     $JAVA_HOME/bin/native-image Streams

    An executable file, streams, is created in the current working directory. Now run it with the same arguments to see the performance:

     ./streams 100000 200

    This version of the program is expected to run slower than on GraalVM’s or any regular JDK.

  4. Build an instrumented native executable by passing the --pgo-instrument option to native-image:

     $JAVA_HOME/bin/native-image --pgo-instrument Streams

  5. Run it to collect the code-execution-frequency profiles:

     ./streams 100000 20

    Notice that you can profile with a much smaller data size. Profiles collected from this run are stored by default in the default.iprof file.

    Note: You can specify where to collect the profiles when running an instrumented native executable by passing the -XX:ProfilesDumpFile=YourFileName option at run time.

  6. Finally, build an optimized native executable by specifying the path to the collected profiles:

     $JAVA_HOME/bin/native-image --pgo=default.iprof Streams

    Note: You can also collect multiple profile files, by specifying different filenames, and pass them to the native-image tool at build time.

    Run this optimized native executable timing the execution to see the system resources and CPU usage:

     time ./streams 100000 200

    You should get the performance comparable to, or faster, than the Java version of the program. For example, on a machine with 16 GB of memory and 8 cores, the TOTAL time for 10 iterations reduced from ~2200 to ~270 milliseconds.

This guide showed how you can optimize native executables for additional performance gain and higher throughput. Oracle GraalVM offers extra benefits for building native executables, such as Profile-Guided Optimizations (PGO). With PGO you “train” your application for specific workloads and significantly improve the performance.