Restrictions for lock coarsening?
Thomas.Rodriguez at Sun.COM
Tue Jan 6 15:51:57 PST 2009
On Jan 6, 2009, at 11:54 AM, Clemens Eisserer wrote:
>> It's definitely an unsatisfying solution.
> Well, at least in my case it is - eliminating the possibility to
> benefit of biased locking at all :-/
> Are there ideas how the current behaviour could be changed to remove
> that restriction while not negativly influcencing startup?
I think we'd need to revisit the reason for the startup delay in the
first place. It's possible that other changes in biased locking have
reduced or eliminated the penalty that the startup delay attempts to
work around. No one is currently looking at it and I don't believe
our startup benchmarks are open source so external developers can't
really run them to investigate it. So it's a warm body problem
tangled with some process problems.
Maybe, as I think Christian was implying, some mechanism to take
unbiasable objects that turn out to be accessed by a single thread and
convert them into biasable objects would be a good technical
solution. The basic machinery of biased locking is very much complete
but there a still some unanswered questions policy wise that might
result in better performance. It would require a lot of investigation
on someone's part though.
More information about the hotspot-runtime-dev