container base image "blank slate" wants and needs

[jmtd]

I'm the maintainer of the Red Hat OpenJDK container images which are used as base images for Quarkus in some modes.

I'm interested in having a discussion to determine what Quarkus wants and needs from a base image. Our RH images are trying to support other users as well, which means they have more features and baggage than you might want.

We might have an opportunity to re-design things in the near future and Quarkus is a very important downstream user, so I'd love to gather what works, and what doesn't.

Thank you!

[cescoffier]

Hello,

Perfect timing. @mschorsch proposed switching the GC to G1 in #47764. I think it deserves a discussion to see if we could influence that.

[kittylyst]

My understanding is that the short answer here is: It Depends. Specifically, on the number of CPUs available to the container.

We obviously need to do a proper benchmarking exercise with a (range of, if at all possible) non-trivial payload.

But the currently understood wisdom is that Parallel is probably the best choice in a "container that appears to the JVM to be a single core machine" and G1 is the best default for > 1 apparent CPU

[geoand]

For GC in containers, https://learn.microsoft.com/en-us/azure/developer/java/containers/overview has some good general information

[gsmet]

But the currently understood wisdom is that Parallel is probably the best choice in a "container that appears to the JVM to be a single core machine" and G1 is the best default for > 1 apparent CPU

Yeah so that would probably make G1 a better choice for most workloads.
Or is it something you could include in the run script? And choose a wise default if not explicitly defined?

[gsmet]

Thanks for starting this conversation.

There has been discussion that the default JAVA_MAX_MEM_RATIO of 80 might be too high in the Quarkus case as Netty also allocates native memory and it might require more than 20% of the RAM.

Now I suppose it's hard to find a value that matches everyone's usage, but for anything using Netty, 80 might be slightly too high.

That's probably something we should tweak in our default Dockerfiles and document properly.