On 07/08/2014 03:00 PM, Ivan Gerasimov wrote:
I took your latest version of the patch and modified it a little:
http://cr.openjdk.java.net/~plevart/jdk9-dev/IdentityHashMap/webrev.01/
But isn't it post-insert-resize vs pre-insert-resize problem Doug
mentioned above?
I've tested a similar fix and it showed slow down of the put() operation.
Hi Ivan,
Might be that it has to do with # of bytecodes in the method and
in-lining threshold. I modified it once more, to make put() method as
short as possible:
http://cr.openjdk.java.net/~plevart/jdk9-dev/IdentityHashMap/webrev.05/
With this, I ran the following JMH benchmark:
@State(Scope.Thread)
public class IHMBench {
Map<Object, Object> map = new IdentityHashMap<Object, Object>();
@Benchmark
public void putNewObject(Blackhole bh) {
Object o = new Object();
bh.consume(map.put(o, o));
if (map.size() > 100000) {
map = new IdentityHashMap<Object, Object>();
}
}
}
I get the following results on my i7/Linux using:
java -Xmx4G -Xms4G -XX:+UseParallelGC -jar benchmarks.jar -f 0 -i 10 -wi
8 -gc 1 -t 1
Original:
Benchmark Mode Samples Score Score
error Units
j.t.IHMBench.putNewObject thrpt 10 13088296.198 403446.449
ops/s
Patched:
Benchmark Mode Samples Score Score
error Units
j.t.IHMBench.putNewObject thrpt 10 13180594.537 282047.154
ops/s
Can you run your test with webrev.05 and see what you get ?
Regards, Peter
