High Reference Processing/Object Copy time

[Thomas Schatzl]

Hi,

On Thu, 2017-07-13 at 08:51 -0700, Sundara Mohan M wrote:
> Thanks for tuning link.
> 
> I have 2 more JVM (31G heap with CMS GC) running on same machine but
> gc logs never showed sys time > 1.4 only the JVM running with G1GC
> shows very high sys time.
> Another thing noticed is resident size is high compared to another
> process which has CMS GC.
>   PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
>  1438 test        20   0 42.0g  27g  34m S 283.8 21.7   5232:24 java
> . >>> CMSGC
>   619 test         20   0 42.0g  26g  34m S 105.0 21.3   5085:47 java
> . >>> CMSGC
>   600 test        20   0 43.4g  33g  33m S  4.3 26.9 361:54.17 java .
>  >>> G1GC
> 
> 
> Is this something to do with G1GC region based allocation?

Yes, regions play into that, most likely remembered sets (used to do
per-region evacuation, but can't tell without an NMT snapshots [0], and
even then it is somewhat guessing).

You can decrease remembered set memory usage by increasing heap region
size (I think it is only 16m in your case, i.e. you are free to try
32m), or try to rebalance remembered set memory usage.

The threads [1] and [2] give some insights on what to look for and what
knobs to turn.

JDK9 contains some improvements in the remembered set area.

Also make sure to compare apples-to-apples. G1 has different heap
sizing mechanisms, so it might just use a larger part of your provided
Java heap memory. NMT can also show you that in addition to your logs.

Thanks,
  Thomas

[0] https://docs.oracle.com/javase/8/docs/technotes/guides/troubleshoot
/tooldescr007.html
[1] http://mail.openjdk.java.net/pipermail/hotspot-gc-use/2017-February
/002609.html
[2] http://mail.openjdk.java.net/pipermail/hotspot-gc-use/2
017-March/002640.html