G1 parallel old gen collection

Jenny Zhang yu.zhang at oracle.com
Tue Oct 4 20:11:43 UTC 2016


On 10/4/2016 1:01 PM, Vitaly Davidovich wrote:
>
>
> On Tue, Oct 4, 2016 at 3:37 PM, Christian Thalinger 
> <cthalinger at twitter.com <mailto:cthalinger at twitter.com>> wrote:
>
>     First, I apologize if I get all the terminology wrong.  You know
>     me… this is all a blackbox to me.
>
>     I’m currently trying to run some Twitter services with G1 and the
>     results are encouraging but not good enough.  One issue is that
>     full GCs take much longer than with ParallelOld (3-4x longer).
>
> As Charlie says, you really want to avoid Full GCs in G1 - they're 
> super slow, and are really a last resort. They're not intended to be 
> part of normal operation.
>
> I've seen them occur when concurrent marking is either started too 
> late or concurrent marking loses the race with java mutators.  There 
> are several ways to mitigate this, but as Charlie said, you'll need to 
> provide more info Chris.
Also with jdk9, the ihop is dynamic adjusted to avoid the 'started too 
late'.

Jenny
>
>
>     Tony said that G1’s old collection is not parallel.
>
>     Is that still the case in JDK 9?  If yes, is someone at Oracle
>     working on parallelizing it?
>
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.java.net/pipermail/hotspot-gc-dev/attachments/20161004/bd9e9aa1/attachment.htm>


More information about the hotspot-gc-dev mailing list