Ergonomics Profile

[brunoborges]

Initial implementation of the Ergonomics Profile proposal (see JEP Draft).

There are two major things missing from this proof of concept:

1. What to do with Client/Server Class Machines algorithm?
2. Tests

[karianna]

Also needs a review from VM Eng.

[d3r3kk]

@brunoborges please protect this from merging into our main - which is a mirror only.

We may want to merge to an ms-patches branch or some such, but not main.

[brunoborges]

Not my intention to merge. Just to see the difference.

[brunoborges]

I'll convert this to a Draft PR just to be sure :-)

[jerboaa]

Hi Bruno, a couple of comments. I'm a bit concerned about the additional complexity about GC selection. Also keep in mind that those MaxRAMPercentage changes seem to affect a wider audience than intended.

[brunoborges]

Hey Severin, just taking note of your suggestion about disabling UseDynamicNumberOfCompilerThreads for small containers. As I mentioned over email, I think it can make sense, but for small containers up to 512M.

I wonder what others stumbling upon this would think.

[karianna]

My nitpick review is done but other folks who understand the semantics of this change better should review as well :-)

meant to add comment, not approve

[jerboaa]

Hey Severin, just taking note of your suggestion about disabling UseDynamicNumberOfCompilerThreads for small containers. As I mentioned over email, I think it can make sense, but for small containers up to 512M.

This idea came about with the discussions in https://bugs.openjdk.org/browse/JDK-8296125 and https://git.openjdk.org/jdk/pull/10918

[brunoborges]

What to do with MaxDirectMemorySize ?

[brunoborges]

What to do with ReservedCodeCacheSize? By default it is 240M and it doesn't matter how much memory is set in the environment.

InitialCodeCacheSize is 2M.

[brunoborges]

@lifey is suggesting a few changes: microsoft/openjdk-proposals@main...lifey:openjdk-proposals:main

[karianna]

@lifey is suggesting a few changes: microsoft/[email protected]:openjdk-proposals:main

Recommend @mo-beck and Kirk review that diff

[schlosna]

The proposed changes broadly seem reasonable given many of the applications and environments I have seen throughout the years, and simplifies some of the challenges folks have when moving JVM applications into hard resource limited containers.

Do you have thoughts how this might evolve over time? How should folks think about using (ideally just the auto) ergonomics profile, and what is the threshold at which additional JVM argument tuning needs to be considered?

[brunoborges]

The auto profile will be removed and instead I will be adding a flag called -XX:+AutoErgonomicsProfile, while the other flag remains to select between valid profiles. "Auto" is not meant to be a profile, and thus I am thinking about this approach.

The JEP text has changed already to reflect this. Mind to take a look at that?

Regarding more profiles in the future, it is not a goal of this JEP to consider that possibility for OpenJDK, but specific JDK vendors may choose to add more if they want to.