deoptimization insanity

[Alen Vrečko]

Hello everyone.

I am looking at a case where around 30% of CPU time is spent by the JVM in deoptimizations (done the profile with async-profiler).

I can locally reproduce the case. The log just explodes with deoptimization messages. A few methods get deoptimized all the time. I didn't stopwatch the deoptimization log count but it is in the order 100s-1000s of them in a span of few seconds - consistently.

Did JIT eliminate a branch by mistake? Then it figured it should be there?

If there is a high number of method calls, could it trigger some dead code elimination prematurely?

Any ideas?

Thank you

Alen

[Andrei Pangin]

Hi Alen,

If the same methods keep getting deoptimized again and again, this can be a JVM bug. I saw this happening several times:

 1. During static initialization: https://pangin.pro/posts/computation-in-static-initializer
 2. https://stackoverflow.com/a/66019887/3448419

If JVM runs with JDWP agent, deoptimization may also happen every time an exception is thrown.

API like MutableCallSite.setTarget can also provoke deoptimization.

So, there can be many reasons. Try to find the cause with -XX:+LogCompilation or with the help of async-profiler as shown in the above SO answer.

Regards,
Andrei

понедельник, 19 декабря 2022 г. в 18:56:30 UTC, alen....@gmail.com:

[Martijn Verburg]

Just saw this super late - sorry!  Which version of the JVM were you using and do you have a small reproducer?

--
You received this message because you are subscribed to the Google Groups "mechanical-sympathy" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mechanical-symp...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/mechanical-sympathy/CACOdjgQaGqHN_uPU_EjowQ5_7jBfKE0a5kLeWSxScE9OTBdhRg%40mail.gmail.com.

--

Cheers, Martijn (Sent from Gmail Mobile)

[Alen Vrečko]

I apologize for not following up on this sooner.

Andrei, spot on with the deoptimizations on JDWP agent. This JDWP deopt insanity is what I am observing in my development environment.

Martijn, as I was chasing JDWP deopts I refactored the code a little bit and since then not seeing any deoptimization issues on production hardware when running performance tests. 

Now I am seeing performance improvements consistent across all of the performance tests. Looks very good. Before some performance tests performed worse with the optimized code which confused me quite a lot.

A bit unsure which code change fixed it. But don't feel like spending more time on this for now as the problem is fixed.

This has been an interesting learning experience.

Thank you

Alen

V V sob., 24. dec. 2022 ob 07:28 je oseba Martijn Verburg <martijn...@gmail.com> napisala:

To view this discussion on the web, visit https://groups.google.com/d/msgid/mechanical-sympathy/CAP7YuASP_f264NwsrCc3WLNGOhCg3Kn6J2hH-8d_bM9FSamUng%40mail.gmail.com.