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

[Hamlin Li]

On Fri, 7 Jul 2023 16:02:57 GMT, Sergey Kuksenko <skuksenko at openjdk.org> wrote:

>> Hi,
>> I'm not sure if I understand this improvement correctly.
>> I'm not quite familiar with JMH and it's annotations, but seems to me, the change from `@State(Scope.Benchmark)` to `@State(Scope.Thread)` should not improve the performance by reducing cache contention, as in the jmh doc it says "State objects are usually injected into Benchmark methods as ***arguments***, and JMH takes care of their instantiation and sharing.", this seems mean that @State only matters when the annotated class is used as a parameter of a @Benchmark method, but in the tests you modifed, seems there is no such use case.
>> Please also check the sample usages at https://github.com/openjdk/jmh/blob/master/jmh-samples/src/main/java/org/openjdk/jmh/samples/JMHSample_03_States.java.
>
> @Hamlin-Li 
> The PR is fully correct.
> Don't forget, every Java instance method has a specific argument which called "this".  That is why @State annotation is working.

@kuksenko @swati-sha Thanks for explanation. I can understand what you said.
But I'm still not quite sure, as I remember jmh does some code manipulation or instrumentation at source code (or bytecode level?), so the jmh test code you write or see might not be the exact code to be executed at runtime.
It's better to be reviewed further by some one more familiar with jmh, or could you add some data comparing the performance difference, so we can tell it easily?

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

PR Comment: https://git.openjdk.org/jdk/pull/14746#issuecomment-1628469975