RFR: 8311178: JMH tests don't scale well when sharing output buffers

[John Jiang]

On Sat, 1 Jul 2023 07:53:17 GMT, Swati Sharma <duke at openjdk.org> wrote:

> The below benchmark files have scaling issues due to cache contention and leads to poor scaling when run on multiple threads. The patch sets the scope from benchmark level to thread level to fix the issue:
> - org/openjdk/bench/java/io/DataOutputStreamTest.java
> - org/openjdk/bench/java/lang/ArrayCopyObject.java
> - org/openjdk/bench/java/lang/ArrayFiddle.java
> - org/openjdk/bench/java/time/format/DateTimeFormatterBench.java
> - org/openjdk/bench/jdk/incubator/vector/IndexInRangeBenchmark.java
> - org/openjdk/bench/jdk/incubator/vector/MemorySegmentVectorAccess.java
> - org/openjdk/bench/jdk/incubator/vector/StoreMaskedBenchmark.java
> - org/openjdk/bench/jdk/incubator/vector/StoreMaskedIOOBEBenchmark.java
> - org/openjdk/bench/vm/compiler/ArrayFill.java
> - org/openjdk/bench/vm/compiler/IndexVector.java
> 
> Also removing the static scope for variables in org/openjdk/bench/jdk/incubator/vector/VectorFPtoIntCastOperations.java for better scaling.
> 
> Please review and share your feedback.
> 
> Thanks,
> Swati

Not review this PR, but just raise a question.
Should a JMH test, at least in JDK repo, always uses `@State(Scope.Thread)`, even though it uses only one thread?

I just looked through those JMH tests, and found all of them, like the bellows, don't specify the number of threads via `@Threads`.


org/openjdk/bench/java/io/DataOutputStreamTest.java
org/openjdk/bench/java/lang/ArrayCopyObject.java


I suppose the default number of threads is 1.
Maybe the default value will be overridden via the commands when running these JMH tests in bulk (?)

-------------

PR Comment: https://git.openjdk.org/jdk/pull/14746#issuecomment-1629261727