RFR: 8319048: Monitor deflation unlink phase prolongs time to safepoint

[David Holmes]

On Tue, 7 Nov 2023 20:18:07 GMT, Y. Srinivas Ramakrishna <ysr at openjdk.org> wrote:

>>>  Is there anything special happening during those benchmark runs?
>> 
>> I have no way to answer that. The benchmarks are run in our performance measurement system, which is just a black-box to me. I tell it what benchmarks to run, and point it to the two binaries and then I get a bunch of results. FWIW I see for baseline versus patched:
>> 
>> 77.73± 1.65 | 122.95± 44.39  p = 0.001
>> 
>> 
>> but the results you pasted, if I am reading them right, indicate the baseline is ~2x better than patched: 2ns/op versus 4ns/op. ???
>
>> > Is there anything special happening during those benchmark runs?
>> 
>> I have no way to answer that. The benchmarks are run in our performance measurement system, which is just a black-box to me. I tell it what benchmarks to run, and point it to the two binaries and then I get a bunch of results. FWIW I see for baseline versus patched:
>> 
>> ```
>> 77.73± 1.65 | 122.95± 44.39  p = 0.001
>> ```
>> 
> 
> @dholmes-ora : Which benchmark are these numbers with (and on which platform/os/isa)? Additional info from `-Xlog:monitorinflation` (on baseline & specimen), as @shipilev indicated, might shed some more light.

@ysramakrishna as above openjdk.bench.java.lang.invoke.MethodHandleInvokeWithArgs.testInvoke_WithArguments macos-Aarch64

But I've been told by our perf folk that the initial set of results I got would need to be re-run and re-examined in a lot more detail to see if there are really any regression there. In all all far too complex and time-consuming so I will just wait for the usual end-of-release benchmarking to tell us whether there are any concerns.

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

PR Comment: https://git.openjdk.org/jdk/pull/16412#issuecomment-1801076695