Virtual thread overhead in SharedRuntime::handle_wrong_method

[Ron Pressler]

Thank you for reporting this. This may well be a bug. We’re looking into it.

— Ron

> On 21 Aug 2023, at 10:18, Gillespie, Oli <ogillesp at amazon.co.uk> wrote:
> 
> Hi,
> 
> I've been using virtual threads and I noticed some apparently significant overhead in SharedRuntime::handle_wrong_method.
> 
> A simple 'benchmark'/stress test profiled with async-profiler shows (among other similar stacks):
> 
> ```
> --- 29055613704 ns (21.14%), 288804 samples
>  [ 0] ImmutableOopMap::update_register_map(frame const*, RegisterMap*) const
>  [ 1] vframeStream::vframeStream(JavaThread*, bool, bool, bool)
>  [ 2] SharedRuntime::find_callee_method(JavaThread*)
>  [ 3] SharedRuntime::reresolve_call_site(JavaThread*)
>  [ 4] SharedRuntime::handle_wrong_method(JavaThread*)
>  [ 5] jdk.internal.vm.Continuation.enterSpecial
>  [ 6] jdk.internal.vm.Continuation.run
>  [ 7] java.lang.VirtualThread.runContinuation
>  [ 8] java.lang.VirtualThread$$Lambda.0x00007ffad006e3d8.run
>  [ 9] java.util.concurrent.ForkJoinTask$RunnableExecuteAction.exec
>  [10] java.util.concurrent.ForkJoinTask.doExec
>  [11] java.util.concurrent.ForkJoinPool$WorkQueue.topLevelExec
>  [12] java.util.concurrent.ForkJoinPool.scan
>  [13] java.util.concurrent.ForkJoinPool.runWorker
>  [14] java.util.concurrent.ForkJoinWorkerThread.run
> ```
> 
> Is this just the expected path+cost of swapping between virtual threads, or is there something surprising happening here?
> 
> My code for the test is:
> 
> ```
> import java.util.concurrent.*;
> 
> public class Virt {
>    public static void main(String[] args) throws Exception {
>        while (true) {
>            ExecutorService e = Executors.newVirtualThreadPerTaskExecutor();
>            for (int i = 0; i < 100_000; i++) e.submit(Main::task);
>            e.shutdown(); e.awaitTermination(100, TimeUnit.MINUTES);
>        }
>    }
>    static void task() { }
> }
> ```
> 
> Notes:
> - I'm running with JDK21, using `java -XX:+UnlockExperimentalVMOptions -XX:-DoJVMTIVirtualThreadTransitions Virt.java` since I know async-profiler as a JVMTI agent triggers some other overheads which would be distracting in the profiling data.
> - Another user saw similar behaviour in https://github.com/async-profiler/async-profiler/issues/779#issuecomment-1651225252.
> 
> Apologies if this is just totally expected behaviour, but I couldn't find any discussion on it.
> 
> Regards,
> 
> Oli
> 
> 
> 
> Amazon Development Centre (London) Ltd. Registered in England and Wales with registration number 04543232 with its registered office at 1 Principal Place, Worship Street, London EC2A 2FA, United Kingdom.
> 
> 
>