Re: VolanoMark performance improvements (was: Re: volanoMark 12% regression with 2.6.25-rc6)

["Zhang, Yanmin" <yanmin_zhang@linux.intel.com>]

On Wed, 2008-03-19 at 14:48 +0100, Ingo Molnar wrote:
> i just tried a handful of kernels with VolanoMark on an Intel quad-core 
> testbox,
How many physical processors does it have?

>  using java-1.6.0-openjdk-1.6.0.0-0.4.b06.fc9.
> 
> The result: v2.6.25 is the fastest ever Linux kernel in VolanoMark 
> testing, -rc7-to-be is ~41% faster than v2.6.24.2, and ~59% faster than 
> v2.6.21, in this Java benchmark:
Did you set /proc/sys/kernel/sched_compat_yield=1 before testing? If not,
pls. configure it.

> 
> v2.6.25-rc7-to-be:
> 
>   Average throughput = 145613 messages per second
>   Average throughput = 142569 messages per second
>   Average throughput = 144963 messages per second
>   Average throughput = 144998 messages per second
>   Average throughput = 144213 messages per second
>   Average throughput = 144491 messages per second
> 
> v2.6.24.2:
> 
>   Average throughput = 101351 messages per second
>   Average throughput =  99247 messages per second
>   Average throughput = 101155 messages per second
>   Average throughput = 103905 messages per second
>   Average throughput =  98345 messages per second
>   Average throughput = 102775 messages per second
> 
> v2.6.21:
> 
>   Average throughput =  87630 messages per second
>   Average throughput =  91366 messages per second
>   Average throughput =  91317 messages per second
>   Average throughput =  91317 messages per second
>   Average throughput =  91359 messages per second
>   Average throughput =  92116 messages per second
> 
> these numbers are totally reproducible when running them over a long 
> time, i only included 6 runs for brevity. (A sidenote: the 
> /proc/sys/kernel/compat_sched_yield switch of 0 or 1 has no effect on 
> the .25-rc7 results, and it degraded the .24 results by about 35% so i 
> kept it off there.)
> 
> so i can see no 12% regression at all - but i have not tried all the 
> zillions of older, binary-only Java distributions.
> 
> So, to make the tests comparable, and to make some progress on this 
> matter, could you please try java-1.6.0-openjdk - what kind of 
> VolanoMark numbers do you get with it?
On my stoakley (2 quad-core processors, 2.66GHz), with jrockit-R27.3.1-jre1.5.0_11.x86_64:
v2.6.22:363062
v2.6.24:390248
v2.6.25-rc3:210276
v2.6.25-rc4:379704
v2.6.25-rc5:377723
v2.6.25-rc6:322864
The regression is reproducible.


On another tigerton(4 quad-core processors, 2.93GHz):
v2.6.25-rc5:577752
v2.6.25-rc6:579936
So there is no regression on tigerton.


>  If you still see a regression 
> with an open Java package too then could you please send me the .config 
> of your test-kernel? 
I will try openjdk.

-yanmin