RFR(XL): 8203469: Faster safepoints

[Robbin Ehn]

On 2019-01-15 11:48, Andrew Haley wrote:
> On 1/15/19 10:39 AM, Robbin Ehn wrote:
>> - On a 16 strand machine synchronization and un-synchronization/starting is at
>>     least 3x faster (in non-trivial test). Synchronization ~600 -> ~100us and
>>     starting ~400->~100us.
> 
> Thanks. Could you share that benchmark? It'd be interesting to try it on
> other architectures.
> 

Those numbers are calculate from linux perf data, I couldn't remember what I run.
But apparently it is also specJVM2008 serial with benchmark thread 16.
The reason for serial that is that Aleksey tested a couple of previews with
Shenandoah and indicated that serial was a good test.
(Aleksey, if you have time please test again)
And from my testing he seems to be correct.
Note that the 16 strands is one package.

So we have no logging or time measurement for 'unsynchronization' since it would
need some extra synchronization overhead.

Thanks for testing it!

/Robbin