On Sat, Sep 19, 2009 at 7:26 AM, Ximin Luo <xl269 at cam.ac.uk> wrote:
> I created a new branch (bigint7) which tests the performance of java 7's
> BigInteger implementation. can people please benchmark it?
>
> On latest fred-staging:
>
> $ git checkout bigint7
> $ ant unit -Dbenchmark=true
>
> ? ?[junit] ------------- Standard Output ---------------
> ? ?[junit] DEBUG: Warming up the random number generator...
> ? ?[junit] DEBUG: Random number generator warmed up
> ? ?[junit] INFO: 200 runs complete without any errors
> ? ?[junit] native run time: ? ?2998ms (14ms each)
> ? ?[junit] java run time: ? ? ?9080ms (45ms each)
> ? ?[junit] java7 run time: ? ? 9118ms (45ms each)
> ? ?[junit] native = 33.01762114537445% of pure java time
> ? ?[junit] java 7 = 100.41850220264317% of pure java time
> ? ?[junit] DEBUG: Warming up the random number generator...
> ? ?[junit] DEBUG: Random number generator warmed up
> ? ?[junit] INFO: 200 runs complete without any errors
> ? ?[junit] native run time: ? ?5994ms (29ms each)
> ? ?[junit] java run time: ? ? ?5093ms (25ms each)
> ? ?[junit] java7 run time: ? ? 4498ms (22ms each)


Native is slower then pure java here? flawed benchmark?
Or just some very good jit?

> ? ?[junit] native = 117.6909483604948% of pure java time
> ? ?[junit] java 7 = 88.31729825250343% of pure java time
> ? ?[junit] ------------- ---------------- ---------------
>
> First test is modPow(), second test is doubleValue()
>
> If these tests are typical, it looks like we should remove the native
> implementation of doubleValue, but it doesn't seem that important.
>
> X

Reply via email to