RFR (S): JEP-142: Reduce Cache Contention on Specified Fields

[Aleksey Shipilev]

On 11/23/2012 01:33 AM, Aleksey Shipilev wrote:
> Hi,
> 
> After some internal discussions with Doug Lea, Dave Dice and others, I
> would like to solicit the initial feedback on the implementation of
> JEP-142, aka @Contended [1]:
>   http://openjdk.java.net/jeps/142
> 
> The webrev for the initial version is here:
>   http://shipilev.net/pub/jdk/hotspot/contended/webrev-2/

BTW, here's the simple microbenchmark test: four threads T1..T4 on my
2x2 i5-2520M each increments its distinct field i1..i4:

    static class Base {
        @Contended long i1;
        @Contended long i2;
        @Contended long i3;
        @Contended long i4;
    }

Running with proper warmups, 10 measurement iterations for 1 second
each, 10 JVM invocations per test, yields:

 -XX:FieldPaddingWidth=0:   7.94 +- 1.10 nsec/op
 -XX:FieldPaddingWidth=8:   4.92 +- 0.53 nsec/op
 -XX:FieldPaddingWidth=16:  4.67 +- 0.54 nsec/op
 -XX:FieldPaddingWidth=32:  4.67 +- 0.31 nsec/op
 -XX:FieldPaddingWidth=64:  3.55 +- 0.03 nsec/op
 -XX:FieldPaddingWidth=128: 3.54 +- 0.03 nsec/op

(The time is severely bloated because of hyperthreading and other infra
overheads).

So that's at least two-fold difference even within single CPU package,
where false sharing misses are being served by on-chip L3. It will get
dramatically worse on multi-CPU hosts. "width=0" corresponds to our
current behavior, i.e. dense packing. Note also the jitter is
significantly better since we are not at the mercy of execution
interleavings.

Thanks,
-Aleksey.