RFR: 8307958: Metaspace verification is slow causing extreme class unloading times [v2]

[Thomas Stuefe]

On Tue, 23 May 2023 09:17:26 GMT, Axel Boldt-Christmas <aboldtch at openjdk.org> wrote:

> > A regression test would be good, since Metaspace will evolve and so will verifications, and I would like to know what performance costs are acceptable.
> 
> Oracles internal testing will most likely provoke regressions in this area. Similarly how this issue was brought to light. To see the really degenerate behaviour, required constant loading and unloading of many class loaders with very few classes over hours with infrequent class unloading (due to generational ZGC not performing many major collections, low old heap pressure).
> 
> While it is probably possible to reduce the variables and create a simpler test which solely stresses metaspace debug timings and little else, it seems like a rather complex task.
> 
> Any product performance regression is caught by our normal performance testing and any relevant and degenerate debug regression would be caught by Oracle internal tests. Hopefully this is enough for now?

Sure, if you guys are okay with defusing my future too-eager verifications :-) ?

I'd be happy with something simple and pragmatic. It does not have to be automatic. Tell me a text X and a time Y it should not overstep on reasonable hardware. Can be very fuzzy. Then I can test this myself instead of causing work for you.

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

PR Comment: https://git.openjdk.org/jdk/pull/14084#issuecomment-1558904152