To Cache or Not to Cache Virtual Threads

[giscus[bot]]

To Cache or Not to Cache Virtual Threads

Quarkus: Supersonic Subatomic Java

https://quarkus.io/blog/to-cache-or-not-to-cache-virtual-threads/

[vsevel]

is it possible to reproduce this benchmark in our environment? is the source code of the 2 quarkus and SB apps public?

[franz1981]

The benchmark is based netty running as VirtualThread with a new experimental feature for loom we are working on, see openjdk/loom#226

The benchmark is at https://github.com/franz1981/Netty-VirtualThread-Scheduler/tree/virtual-netty-loop-experiment/benchmark-runner but it doesn't contains the "virtual thread pooled executor" in the article; I still have to push it.

And the jdk patch to prevent stack chunks to be allocated is at openjdk/loom#228

Which mean you need both to compile JDK and the mentioned benchmark to fully reproduce the findings.

My suggestion is to just modify Quarkus to pool VTs, compile it and use it on https://github.com/quarkusio/spring-quarkus-perf-comparison
in case you want to just observe the pooling in action without any JDK patch, but you need to have enough heap to prevent the pooled VTs to age too aggressively and get their StackChunk being allocated fresh new causing the churning I have explained in https://quarkus.io/blog/to-cache-or-not-to-cache-virtual-threads/#what-could-go-wrong

I expect that pooling still have some value assuming you have enough heap to prevent too frequently these stack chunks to be left to old gen and freshly allocated

[vsevel]

I commented on the wrong article ;-)
I meant this benchmark : https://quarkus.io/blog/fairness-in-benchmarking/

[franz1981]

I can setup a call once I am back from vacation for that one (25 Aug)

[holly-cummins]

I commented on the wrong article ;-) I meant this benchmark : https://quarkus.io/blog/fairness-in-benchmarking/

I'm pretty sure that one is https://github.com/quarkusio/spring-quarkus-perf-comparison/ (see quarkusio/spring-quarkus-perf-comparison#633). There's full instructions for reproducing in the readme.