JFR sampling question

[Erik Gahlin]

Hi Thomas,

> On 4 Nov 2020, at 17:26, Thomas Stüfe <thomas.stuefe at gmail.com> wrote:
> 
> Hi Eric,
> 
> may I ask some more questions?
> 
> 1) Via experiments I found that in JMC, the graphs under "JVM Internals->Garbage Collections" which describe Metaspace development (reserved, committed, used) seem to be calculated from the MetaspaceSummary Event. Is that correct?

Sorry. I don’t know what the GUI shows. 

You may want to check with jmc-dev at openjdk.java.net, or the JMC source code:

https://github.com/openjdk/jmc/tree/master/application/org.openjdk.jmc.flightrecorder.ui/src/main/java/org/openjdk/jmc/flightrecorder/ui/pages

> 
> 2) These graphs seem to be interpolated, right? Since regardless of how many events I actually have, they look smooth. Is there a way to show the actual samples instead of the graphs, or maybe the sample points interposed on the line?

You need to check with JMC team.

> 
> 3) I did not find any specification for the jfc files, is there one?

See:

src/jdk.jfr/share/classes/jdk/jfr/internal/jfc/jfc.xsd

> From experimenting and looking at the jfc file, looks like MetaspaceSummary is only sent when a GC happened, is that correct?

I have not implemented the event, but it does a have gcId field which means it is tied to a GC.


> The problem with that is that a lot of information goes missing since metaspace state can change a lot outside of GCs. What would be the correct way to change these events to periodic delivery (if that is even possible)?

Events can be made periodic in which case they are fired at an interval. You would set the “period” attribute, see src/hotspot/share/jfr/metadata, and then emit your event in src/hotspot/share/jfr/periodic/jfrPeriodic.cpp

Perhaps there should be two events, one that happens every GC and one that is emitted periodically? The periodic one can’t have a gcId.

There is some precedence for creating ordinary events and statistical events where the latter aggregate data over time.  Not sure if it applies in your case? There is also events like ObjectCount and ObjectCountAfterGC that looks identical but one is emitted periodically and the other after a GC.

Do you have any numbers of the cost of aggregating data for the event?

You can regulate the overhead by setting the period. “endChunk” which means it is included at least once in every recording, or you can set the interval “1 s”, “10 s”, “30 s” or “60 s” depending on what is needed for troubleshooting and the overhead (space, CPU) the event creates.

I hope this helps.

If you want more guidance, it may be easiest to just set upp 30 min zoom meeting.

Cheers.
Erik

> 
> Thanks a lot!
> ..Thomas
> 
> 
> On Wed, Nov 4, 2020 at 1:27 PM Thomas Stüfe <thomas.stuefe at gmail.com <mailto:thomas.stuefe at gmail.com>> wrote:
> Thank you Erik, that is exactly what I needed.
> 
> On Wed, Nov 4, 2020 at 1:13 PM Erik Gahlin <erik.gahlin at oracle.com <mailto:erik.gahlin at oracle.com>> wrote:
> Hi Thomas,
> 
> I don’t know what the event is called, but you can look at the configuration files and the period setting for the event.
> 
> https://github.com/openjdk/jdk/tree/master/src/jdk.jfr/share/conf/jfr <https://urldefense.com/v3/__https://github.com/openjdk/jdk/tree/master/src/jdk.jfr/share/conf/jfr__;!!GqivPVa7Brio!I66M-vBapUnyoKtZf-uHXqM4jTZkZL9TN8pWFKesEyzfV0MUyl3z6lDdJtvc84Z6$>
> 
> Erik
> 
> > On 4 Nov 2020, at 13:07, Thomas Stüfe <thomas.stuefe at gmail.com <mailto:thomas.stuefe at gmail.com>> wrote:
> > 
> > Hi,
> > 
> > I am currently reworking the Metaspace information shown in JMC, which is
> > outdated after JEP387.
> > 
> > How often is this information sampled? How expensive can the sampling code
> > get?
> > 
> > For example, post-JEP387, the metachunk freelists contain committed as well
> > as uncommitted areas. While it may be interesting to know how much area is
> > committed, that would require walking the freelists, which could get
> > expensive.
> > 
> > Thanks, Thomas
>