CMS Concurrent mode failure fallback to the serial old collector?

[Jon Masamitsu]

On 08/17/12 14:08, Haim Yadid wrote:
> I am analysing a GC pause problem and I have noticed that when CMS is 
> used and a concurrent mode failure occurs or GC is triggered manually 
> (by System.gc()) the STW collector used does not seem to be parallel. 
> ( I am aware of the ExplicitGCInvokesConcurrent flag but it will not 
> solve concurrent failure ).
> I tried to play with -XX:ParallelGCThreads=... 
> -XX:ParallelCMSThreads=... but they seem have no effect (only on the 
> ParNew GC).
>
> I am deducing it from the following GC log line
>
> 24.904: [Full GC (System) 24.904: [CMS: 302703K->303056K(2116864K), 
> 1.0847520 secs] 484492K->303056K(2423552K), [CMS Perm : 
> 7528K->7525K(21248K)], 1.0852780 secs] [Times: user=1.04 sys=0.02, 
> real=1.09 secs]
> If it would have been parallel "user" would have been equal to 
> "nThreads" * "real".
> In addition if I choose ParallelOld GC it will behave correctly.
>
> I really do not understand why the failover STW mechanism of CMS is 
> not parallel shouldn't it be finishing the work as soon as possible ?
> I am not able to find anything useful on the internet.

You are correct that the concurrent mode failure does a full GC serially.
The parallel old collector used for UseParallelGC/UseParallelOldGC was
never ported to CMS.  Because of differences between UseParallelGC and
CMS, it is more work than we had expected.
>
> I think G1 behaves in the same manner BTW ( AFAIK the the fallback 
> collector of G1 is copied from CMS)

Yes, G1 behaves the same.  G1 will not use the UseParallelGC implementation
for a parallel full collection but will implement one in line with G1's 
design.
Currently the G1 guys have been focusing on better policies for achieving
pause goals and avoiding full collections.  Last I heard there was at 
least some
work to be done for class unloading (JEP 156) before the parallel full 
collection.

Jon

>
> Help will be appreciated.
>
>
> _______________________________________________
> hotspot-gc-use mailing list
> hotspot-gc-use at openjdk.java.net
> http://mail.openjdk.java.net/mailman/listinfo/hotspot-gc-use
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/hotspot-gc-use/attachments/20120820/59be1d20/attachment.html