RFR(xs): 8245926: JFR: Remove chunkSz and rename blockSz from classloader statistics

[Thomas Stüfe]

Hi Erik,



On Wed, May 27, 2020 at 10:06 AM Erik Gahlin <erik.gahlin at oracle.com> wrote:

> Hi Thomas,
>
> An important use case for JFR is to have all relevant information
> available for support since they can’t typically run jcmd. If jcmd has
> better information, I think we should strive to make that information
> available for JFR as well.
>
> An alternative would be to create a new event for metaspace targeted for
> JVM engineers/support, with similar information as jcmd(?), and only keep
> fields relevant to a Java developer in ClassLoaderStatistics.
>
> Thanks
> Erik
>
>
Okay, you have a point. I may have misunderstood you before. I'll withdraw
the RFE for now. Since a lot of implementation details will change, if this
kind of detail is to be exposed via JFR, this will need brushing after
JDK-8221173 hits mainline anyway so it makes no sense to put much work into
it now.

My question about backward compatibility remains though: can one just
change/remove event fields? I assume JMC uses those fields for things
beyond displaying events in the generic event browser, yes? How does JMC
handle backward compatibility - is it only supposed to work with one
hotspot or with a range of them?

For instance, the "Class Loader Statistics" Tab in the "Class Loading"
section of the VM still talks about "Anonymous Classes" and shows no
content there; I assume this was broken with the Hidden Classes JEP which
renamed the fields in ClassLoaderStatistics from "Anonymousxxx" to
"Hiddenxxx". So is it bound to one specific hotspot version? Just a guess,
I did not look at the JMC sources.

Thanks, Thomas



> > On 27 May 2020, at 08:57, Thomas Stüfe <thomas.stuefe at gmail.com> wrote:
> >
> > Hi,
> >
> > may I have opinions/reviews for this change please:
> >
> > JBS: https://bugs.openjdk.java.net/browse/JDK-8245926
> > Webrev:
> >
> http://cr.openjdk.java.net/~stuefe/webrevs/8245926-jfr-remove-chunksz-rename-blocksz/webrev.00/webrev/
> >
> > This is part of an effort to reduce exposure of Metaspace implementation
> > details somewhat in preparation for JDK-8221173.
> >
> > In JFR, the classloader statistics event contains the following fields:
> >
> > chunkSize
> > blockSize
> > hiddenChunkSize
> > hiddenBlockSize
> >
> > blockSize is the interesting one, the number of bytes of Metaspace used
> by
> > this class loader for storing metadata. I propose to rename this to
> > "totalMetaspaceUsed" resp. "hiddenClassesMetaspaceUsed".
> >
> > chunkSize is the total size of chunks given to a class loader. Delta
> > between blockSize and chunkSize constitutes waste if the loader stops
> > loading; however, as far as waste goes this is not very informative since
> > it only is a usually small part of waste in the current metaspace
> > allocator.
> >
> > Moreover, with JDK-8221173 chunks may be partially uncommitted, so that
> the
> > delta between chunkSize and blockSize may or may not contribute to RSS.
> So
> > the chunkSize stat becomes more meaningless.
> >
> > I'd argue that this stat is quite useless to the average JFR user; and to
> > the VM developer interested in fixing up Metaspace there are better tools
> > (e.g. jcmd VM.metaspace). Therefore I would like to simply remove this
> stat.
> >
> > --
> >
> > Please note that this is my first change to JFR events. Some things are
> > unclear to me, e.g. how backward compatibility in JMC is handled. So if
> > something is missing from the patch, please advice.
> >
> >
> > Thanks!
> >
> > .. Thomas
>
>