Very long duration of Finalize marking in the Remark stage of G1GC

[Thomas Schatzl]

Hi,

On 01.10.19 03:26, Tony.Wang wrote:
> Hi
> 
> is this problem solved? would you please share it with me?

   I am not completely sure which problem or which thread you are 
answering here though as there is no context.

There has been a thread of that name in 2015 afaict; the last log 
snippet here

2015-10-18T07:07:28.966-0400: 346753.917: [GC remark 
2015-10-18T07:07:28.966-0400: 346753.917: [Finalize Marking, 115.4404545 
secs] 2015-10-18T07:09:24.406-0400: 346869.357: [GC 
ref-proc2015-10-18T07:09:24.406-0400: 346869.357: [SoftReference, 54 
refs, 0.0028579 secs]2015-10-18T07:09:24.409-0400: 346869.360: 
[WeakReference, 418 refs, 0.0020702 secs]2015-10-18T07:09:24.411-0400: 
346869.362: [FinalReference, 1633 refs, 0.0025592 
secs]2015-10-18T07:09:24.414-0400: 346869.365: [PhantomReference, 6 
refs, 21 refs, 0.0040350 secs]2015-10-18T07:09:24.418-0400: 346869.369: 
[JNI Weak Reference, 0.0001088 secs], 0.0130411 secs] 
2015-10-18T07:09:24.419-0400: 346869.370: [Unloading 
2015-10-18T07:09:24.420-0400: 346869.371: [System Dictionary Unloading, 
0.0001458 secs] 2015-10-18T07:09:24.420-0400: 346869.371: [Parallel 
Unloading, 0.0153302 secs] 2015-10-18T07:09:24.435-0400: 346869.386: 
[Deallocate Metadata, 0.0000526 secs], 0.0159394 secs], 115.5038062 secs]

with the long Finalize Mark phase could be fixed for newer VM version 
because since that time significant improvements in that code were made 
(much better parallelization).

If you are experiencing the same problem, it would probably be best to 
start over with the analysis (with new log output as a starting point) 
after so much time has passed.

Another potential problem *could* be what is described by 
https://bugs.openjdk.java.net/browse/JDK-8205353; one could try lowering
G1SATBBufferEnqueueingThresholdPercent to 0 - but please do not add 
random flags without proper analysis of potential causes and looking if 
the problem still occurs with a more recent jdk >= 9 first.

Thanks,
   Thomas

> 
> -- 
> Shunyun Wong
> —————————————————————————
> *it was the best of times, it was the worst of times*
> 
> _______________________________________________
> hotspot-gc-use mailing list
> hotspot-gc-use at openjdk.java.net
> https://mail.openjdk.java.net/mailman/listinfo/hotspot-gc-use
>