G1 Performance [Was: JEP 248: Make G1 the Default Garbage Collector]

[Andrew Haley]

On 07/29/2015 06:55 PM, Doug Lea wrote:
> On 07/29/2015 12:09 PM, Andrew Haley wrote:
>> I am concerned that G1 does not perform well with some important
>> workloads.  In particular, on heavily-loaded systems some
>> multi-threaded programs which generate a lot of garbage run
>> significantly more slowly with G1.  I've been measuring the
>> performance loss and it's about 20-30%, depending on the application.
> 
> The problem is that there is a full memory fence when issuing
> cross-region GC write barriers. Most concurrent producer-consumer
> designs hit this case frequently, and can run almost arbitrarily
> slower, down to the rate of issuing fences (MFENCE/DMB/etc).
> I know that Erik Osterlund (and possibly others) are looking into ways
> of removing this fence. I sure hope they succeed before jdk9 ships!

Indeed.

I'm looking at a 5-6% degradation (measured against the parallel GC)
in performance over a range of benchmarks on the system I measured.

As I understand it, one of the acceptance criteria of a Java release
is that it must not have overall worse performance than the previous
Java release.  As it stands, I don't think that this version of Java
is going to meet that requirement.

Andrew.