A DESCRIPTION OF THE PROBLEM :
In previous versions of Java, ThreadLocalRandom.current().doubles().parallel() was approximately the same speed as new SplittableRandom().doubles().parallel(). However, since Java 17, there is heavy contention on the stream when run in parallel.



REGRESSION : Last worked in version 16

STEPS TO FOLLOW TO REPRODUCE THE PROBLEM :
ThreadLocalRandom.current().doubles(100_000_000).parallel().sum() on Java 16 vs Java 17.

EXPECTED VERSUS ACTUAL BEHAVIOR :
EXPECTED -
Expected to complete in about 150 milliseconds.
ACTUAL -
Takes several seconds, depending on the hardware. On 1-8-2 MacBookPro Intel it took 24 seconds. On my 1-6-2 Ubuntu Intel Server it took 4.4 seconds.

---------- BEGIN SOURCE ----------
import java.util.SplittableRandom;
import java.util.concurrent.ThreadLocalRandom;

public class Regression {
    public static void main(String... args) throws InterruptedException {
        int STREAM_SIZE = 32_000_000;

        long srIntsTime = System.nanoTime();
        try {
            System.out.println("sum ints = " +
                    new SplittableRandom().ints(STREAM_SIZE)
                            .parallel()
                            .sum());
        } finally {
            srIntsTime = System.nanoTime() - srIntsTime;
            System.out.printf("srIntsTime = %dms%n", (srIntsTime / 1_000_000));
        }

        long tlrIntsTime = System.nanoTime();
        try {
            System.out.println("sum ints = " +
                    ThreadLocalRandom.current().ints(STREAM_SIZE)
                            .parallel()
                            .sum());
        } finally {
            tlrIntsTime = System.nanoTime() - tlrIntsTime;
            System.out.printf("tlrIntsTime = %dms%n", (tlrIntsTime / 1_000_000));
        }

        long srLongsTime = System.nanoTime();
        try {
            System.out.println("sum ints = " +
                    new SplittableRandom().longs(STREAM_SIZE)
                            .parallel()
                            .sum());
        } finally {
            srLongsTime = System.nanoTime() - srLongsTime;
            System.out.printf("srLongsTime = %dms%n", (srLongsTime / 1_000_000));
        }

        long tlrLongsTime = System.nanoTime();
        try {
            System.out.println("sum longs = " +
                    ThreadLocalRandom.current().longs(STREAM_SIZE)
                            .parallel()
                            .sum());
        } finally {
            tlrLongsTime = System.nanoTime() - tlrLongsTime;
            System.out.printf("tlrLongsTime = %dms%n", (tlrLongsTime / 1_000_000));
        }

        long srDoublesTime = System.nanoTime();
        try {
            System.out.println("sum doubles = " +
                    new SplittableRandom().doubles(STREAM_SIZE)
                            .parallel()
                            .sum());
        } finally {
            srDoublesTime = System.nanoTime() - srDoublesTime;
            System.out.printf("srDoublesTime = %dms%n", (srDoublesTime / 1_000_000));
        }

        long tlrDoublesTime = System.nanoTime();
        try {
            System.out.println("sum doubles = " +
                    ThreadLocalRandom.current().doubles(STREAM_SIZE)
                            .parallel()
                            .sum());
        } finally {
            tlrDoublesTime = System.nanoTime() - tlrDoublesTime;
            System.out.printf("tlrDoublesTime = %dms%n", (tlrDoublesTime / 1_000_000));
        }
    }
}

---------- END SOURCE ----------

CUSTOMER SUBMITTED WORKAROUND :
Use SplittableRandom.

And I have submitted a PR here: https://github.com/openjdk/jdk/pull/12366

(Not in correct format though)

FREQUENCY : always