Files
Claude 4511cc6e6b parallel-streams: JMH benchmarks and demos for when parallel streams are slower
N x Q model, source splittability, boxing, ordered operations, common-pool
starvation, and parallel-stream correctness bugs. Captured output in output/.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_017ic4b6hCTuFrQT8xKmmMrF
2026-11-17 23:23:21 +05:30

3.1 KiB

parallel-streams

Companion code for the ankurm.com post "Java Parallel Streams: When They're Slower (JMH Benchmarks and the Common Pool)." Module in java-core-examples, the Java-core series. There is deliberately no docs/ folder: the deeper material lives in collapsible sections inside the post itself.

Versions this was built and tested against

Component Version Notes
JDK 25.0.5+8 (Temurin, LTS)
JMH 1.37 Latest GA on Maven Central; released 3 Aug 2023.
JUnit Jupiter 5.11.0
Maven 3.9.11
Host 2 vCPU shared sandbox Default common-pool parallelism is therefore 1 (see output/00). Speed-ups cap near 2x. Read the shape of the numbers, not the third significant digit.

Quickstart

export JAVA_HOME=/path/to/jdk-25
mvn -q package                                              # builds target/benchmarks.jar and runs the tests
java -Djava.util.concurrent.ForkJoinPool.common.parallelism=3 \
     -cp target/classes com.ankurm.parallelstreams.CorrectnessBugsDemo
java -Djava.util.concurrent.ForkJoinPool.common.parallelism=3 \
     -cp target/classes com.ankurm.parallelstreams.CommonPoolStarvationDemo
java -jar target/benchmarks.jar NqBenchmark -f 1 -wi 3 -w 1s -i 5 -r 1s

scripts/run-all.sh regenerates every file in output/ (about 6 minutes on 2 vCPUs). scripts/run.sh <ClassName> runs one demo class ad hoc.

What's in here

File What it shows
NqBenchmark.java N elements x Q units of CPU per element, sequential vs parallel, at three sizes and two costs. Where parallel starts to pay.
SourceBenchmark.java The same work over int[], ArrayList, LinkedList and Stream.iterate: how splittable the source is decides the outcome.
BoxingBenchmark.java A cheap sum over 1,000,000 elements, primitive vs boxed, with -prof gc allocation per call.
OrderedOpsBenchmark.java limit and distinct, ordered vs unordered(), against a sequential baseline.
CommonPoolStarvationDemo.java A neighbour that blocks the common pool slows an unrelated parallel stream 4x; a private ForkJoinPool and virtual threads as the fixes.
CorrectnessBugsDemo.java forEach(list::add), a non-identity reduce identity, a non-associative accumulator, and a racy counter.
ParallelStreamsTest.java Pins the deterministic claims. No timing assertions.
output/00-07 Captured runs of all of the above.

Notes worth knowing before reading the post

  • JMH settings are light (-f 1 -wi 3 -i 5, one second each) to keep the whole capture to minutes. Error bars are reported in the output; the post draws conclusions only from differences clearly larger than them.
  • The demos pin common.parallelism=3 so they behave the same on any machine. The JMH runs do not: they use the default for the host, which is 1 on 2 CPUs.
  • BoxingBenchmark's first draft compared two different stream sources (Stream.iterate for the sequential boxed case, range().boxed() for the parallel one). That is not a fair pair; both now use IntStream.range(..).boxed().

License

MIT - see the repo-wide LICENSE.