AW: Re: G1 GC - pauses much larger than target

[Andreas Müller]

I resend the text of my reply. It was blocked from the mailing list because the graphic was a bit too large. Sorry for that but I could not make it smaller.

Von: Andreas Müller
Gesendet: Freitag, 21. Februar 2014 09:39
An: 'kirtiteja at gmail.com'
Cc: 'hotspot-gc-use at openjdk.java.net'
Betreff: Re: G1 GC - pauses much larger than target

Hi Kirti,


> I am trying out G1 collector for our application. Our application runs with 2GB heap and we expect relatively low latency.

> The pause time target is set to 25ms. There >are much bigger pauses (and unexplained) in order of few 100s of ms.

> This is not a rare occurence and i can see this 15-20 times in 6-7 hours runs.



This conforms to what I have observed in extended tests:

G1's control of GC pause duration is limited to a rather narrow range.

Even in that range, only new gen pauses do follow the pause time target well while "mixed" pauses tend to overshoot with considerable probability.

Find attached a graphic which shows what I mean:

-        New gen pauses (red) do follow the target very well from 150-800 millis

-        With a target below 150 the actual new gen pauses remain flat at 150-180 millis

-        "mixed" pauses (blue) do not follow the target well and some of them will always take 500-700 millis, whatever the target be

-        There are other pauses (remark etc., green) which are short but completely independent of the target value



The range with reasonable control depends on the heap size, the application and the hardware.

I measured the graphic attached on a 6-core Xeon/2GHz server running Java 7u45 on CentOS/Linux with 64 GB RAM and a heap size of -Xms50g -Xmx50g.

(For which the pause durations achieved are not bad at all!)

The application was a synthetic benchmark described here: http://blog.mgm-tp.com/2013/12/benchmarking-g1-and-other-java-7-garbage-collectors/

With the same benchmark but only 10 GB of overall heap size on a Oracle T3 server running Java 7u45 on Solaris/SPARC I got a very similar kind of plot but the range with reasonable pause time control was now 60-180 millis.

Again the pause durations reached were by themselves not bad at all. But the idea of setting a pause time target and expecting it to be followed in a meaningful way is to some extent misleading.



These results on G1's pause time control will be published soon on the blog of the link above.

Best regards
Andreas

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/hotspot-gc-use/attachments/20140221/d39c23d2/attachment.html