RFR: 8216283: Allow shorter method sampling interval than 10 ms

Markus Gronlund markus.gronlund at oracle.com
Sat Jun 1 08:16:17 UTC 2019


Hi Erik,

Looks good.

Thanks
Markus

-----Original Message-----
From: Erik Gahlin 
Sent: den 20 maj 2019 19:47
To: hotspot-jfr-dev at openjdk.java.net
Cc: jmc-dev at openjdk.java.net
Subject: RFR: 8216283: Allow shorter method sampling interval than 10 ms

Hi,

Could I please have a review of a fix that allows method sampling to happen at a higher rate. The current limit of 10 ms exists because method sampling used to occur in the VM period task. Method sampling now happens in a separate thread. This was changed in JDK 9.

The minimum accepted interval is lowered from 10 ms to 1 ms for both native and Java samples, but configuration files have been changed so native sampling interval can not be set lower than 20 ms using JMC Recording Wizard controls. Manual edit of the event setting allows 1 ms.

The reason for this is to avoid excessive number native sample events. 
Purpose of native sampling is to detect "stuck threads". That is, threads that have called into native, but not returned. Most of the native samples will however be threads waiting in native, for example on a socket. Going below 20 ms adds little value, since a stuck thread is typically in native for a longer period of time.

Method sampling in Java on the other hand can be set to 1 ms in a new "mode" called "Extreme". To avoid confusion, I didn't want to use "Maximum" since it means 10 ms in earlier releases.

Webrev:
http://cr.openjdk.java.net/~egahlin/8216283/

Bug:
https://bugs.openjdk.java.net/browse/JDK-8216283

Testing:
Tried to set the interval in the JMC Recording Wizard and verified that the correct interval was set in JVM using:
$ jcmd <pid> JFR.check verbose=true

Thanks
Erik


More information about the jmc-dev mailing list