We have this decent server with J-Heap at 20GB. How ever, I just saw a pause the world detected, I though this could only be possible when the memory almost hit the heap, but in the time range where the event happened the memory is far from 20GB. In fact, the avg pick is always 10GB.
Question
What else can produce this clock drift?
Note: Not using Mozilla
@dave.fogle had put up an article on this.
I have found that these messages don’t appear in the log until I start opening the performance pages, such as the threads diagnostics, and overall CPU/ram usage.
I didn’t see this explantation in the article, so don’t know if that’s expected it not.
Your actual delays are just over the limit. Suggesting the JVM’s pause target is one second. You should set that much lower. I typically set a 100 or even 50 millisecond target.
Share your ignition.conf file…
This one. Yes, expected. Avoid looking at threads, as collecting the details on them disrupts the JVM itself.
Could you expound on this a bit?
The JVM can’t capture thread details while they are running. So there’s a brief pause when collection information about all threads for a thread dump. For a busy system with lots of threads, this is noticeable.
It seems a little suspicious that the events are frequently 10s apart or a multiple of 10s. Is this really 7.9? Is there a 10s gateway timer script or anything like that?
Ok… And dropping the pause length will help as well? How do you tune that out properly?
You should have a line in ignition.conf like so:
wrapper.java.additional.2=-XX:MaxGCPauseMillis=100
The G1GC algorithm will run more often when pauses exceed that threshold, and run less often when pauses are substantially less. It is a heuristic, not a hard rule, but G1GC is pretty good at keeping pauses under the target maximum.
Awesome. Always love learning better ways to tune these systems.
Isn’t 10 seconds the refresh pace on gateway status pages?
Not sure. I think the thread viewer page is 5s. Don’t know about others, but that’s the heaviest one I think.
Sorry for delay.
Thanks for all info there.
Indeed, as some one mentioned. This happened due to checking threads page as shows: