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

[Aleksey Shipilev]

On Fri, 3 Nov 2023 05:58:05 GMT, David Holmes <dholmes at openjdk.org> wrote:

> Benchmarks are still running but I am seeing results all over the shop: lots of green but lots of red. One example openjdk.bench.java.lang.invoke.MethodHandleInvokeWithArgs.testInvoke_WithArguments is 58% slower on macos-Aarch64! I need to get the perf folk to take a look at this.

Hm, cannot reproduce that locally on my M1 Mac:


$ CONF=macosx-aarch64-server-release make images test TEST=micro:MethodHandleInvokeWithArgs.testInvoke_WithArguments TEST_VM_OPTS="-Xmx1g -Xms1g -XX:+UseParallelGC"

# Baseline
MethodHandleInvokeWithArgs.testInvoke_WithArguments  avgt   15  68,771 ? 2,188  ns/op

# Patched
MethodHandleInvokeWithArgs.testInvoke_WithArguments  avgt   15  69,299 ? 4,087  ns/op


This is also very weird, because that test does not churn monitors at all (as far as `-Xlog:monitorinflation` can tell), and so monitor deflation thread is not active. Is there anything special happening during those benchmark runs? For example, dumping the thread stacks? That's the only other path I see to `ObjectSynchronizer::deflate_idle_monitors`.

I merged from current master, in case that matters.

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

PR Comment: https://git.openjdk.org/jdk/pull/16412#issuecomment-1792141603