Safepoint timeout in scimark?

Roman Kennke rkennke at redhat.com
Mon Jul 11 17:52:58 UTC 2016


Hi Vladimir,

> Hi, Roman
> 
> Yes, it is known old problem with counted loops which have expensive
> iterations:
> 
> https://bugs.openjdk.java.net/browse/JDK-5014723

Ok thanks. Wow, that's an old one :-)

> But we most likely will not fix it.

Ok. May I ask why not? This behaviour seems very broken, and is very
easy to reproduce in a popular benchmark. 

Also, any idea why I only started running into it quite recently? I am
having an eye on safepoint pause times since a while, and I'm pretty
sure it hasn't happened to me before, at least by far not as reliably.


> Instead we implemented UseCountedLoopSafepoints flag to solve rare
> cases when customers hit this problem.
> 
> You may hit next bug with UseCountedLoopSafepoints flag:
> 
> https://bugs.openjdk.java.net/browse/JDK-8160287
> 
> I added your comment to it.

Hmm, ok, this is bad. Default behaviour is broken IMO, and the flag to
make it right is broken too ;-)

Let me know if I can help in any way. I might even try to fix it
myself, if we can agree what that would look like.

Cheers,
Roman



More information about the hotspot-compiler-dev mailing list