Could we get rid of the UseCompressedClasses switch altogether, defaulting to const true?

[Thomas Stüfe]

To bottom line the discussion and the findings:

Removing UseCompressedClassPointers would limit us to 3GB Klass encoding
range and ~5 million classes. These numbers are very high, so they only are
of interest when generating classes. Generation also includes cases like
Lambdaforms.

Does that matter? There is no way to have a 100% answer here. My bet would
be that these limits are high enough for all real scenarios and that most
scenarios that run into these limits are probably pathological.

There is an unknown number of cases that may run with millions of classes
but don't know it. These are cases that JDK <= 11 with heaps >32G. In these
cases, CompressedClassPointers would silently default to false. If these
cases upgrade their JDK, they will first be exposed to the
UseCCP<->UseCoops decoupling, and if we hardwire UseCCP to true, have no
ability to go back to having millions of classes. This, however, is a
highly theoretical case. There may be no cases like this.

How should we continue? Should I move forward and open a CSR to start the
deprecation process?

Thanks, Thomas


On Wed, Feb 28, 2024 at 6:00 PM Roman Kennke <rkennke at gmail.com> wrote:

> Hi there,
>
> I searched around in Amazon a little bit and all of the (very few) uses of
> -CCP all come from Apache Spark, which has this interesting code:
>
>
> https://github.com/apache/spark/blob/ea2587f695cf34c8561c91ffc342b38995bdeb43/sql/core/src/test/scala/org/apache/spark/sql/execution/WholeStageCodegenSparkSubmitSuite.scala#L45
>
> Looks like they are doing it to get a particular alignment of array
> elements.
>
> This will be shaken up by my recent push of:
> https://bugs.openjdk.org/browse/JDK-8139457
>
> Roman
>
>
> > On Feb 22, 2024, at 3:34 PM, Thomas Stüfe <thomas.stuefe at gmail.com>
> wrote:
> >
> > Greetings,
> >
> > Are there any realistic scenarios that would require
> UseCompressedClassPointers=false? Or could we make it constant at some
> point or remove the switch altogether?
> >
> > The only reason I can see for UseCompressedClassPointers=off would be if
> the maximum class space size of 3GB is not enough. But how realistic is
> this really? Some numbers:
> >
> > On average, a Klass is ~600-700 bytes. Therefore, 3GB of class space
> would be enough for ~5 million classes. Note that the typical size ratio of
> Klass-to-non-class-metadata is 1:6 .. 1:10, which means that 3 GB class
> space come with ~18..30 GB of non-class metaspace (see [1]). I think these
> are outlandish numbers. Do we need to support these scenarios? Some weird
> class generator scenarios maybe?
> >
> > Other than class space max size not being enough, are there any reasons
> for -UseCompressedClassPointers?
> >
> > Thank you,
> >
> > Thomas
> >
> > [1]
> https://github.com/tstuefe/metaspace-statistics/blob/master/Histogram.svg
> >
> >
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/hotspot-runtime-dev/attachments/20240314/9c20ed65/attachment-0001.htm>