2.6 kernel series have noticeable performance regression

2.6 kernel series have noticeable performance regression

[yangyi]

Hi, All

I did a performance regression test for all the 2.6 kernel series on two
x86 platforms in last two weeks and found that some version kernels do
have noticeable performance regression, some performance indices have
about 20 to 30 percent regression,
 even several indices have over 30 percent fall.

For example, on Toshiba tecra8000 laptop: for 4-bytes packet, network
throughput of 2.6.11 is just about 57.5 percent of 2.6.0, for pipe
bandwidth, the value of 2.6.11 is just about 68 percent of that value of
2.6.0, for mmap latency, the value of 2.6.11 is about 1.8 times as long
as 2.6.0.

On supermicro-6012-p6, for 64-bytes small packet, network throughput of
2.6.11 is about 66.72 percent of 2.6.0, for 0K-size 2 processes context
switch time, 2.6.11 is 31.8 percent bigger than 2.6.0, for pipe latency,
2.6.10 is 35.85 percent bigger than 2.6.3.

So, I think that it is very very necessary to find the problems and fix
them ASAP.

The following articles give out all the test results, the article "2.6
kernel series have noticeable performance regression, part 1" gives out
the test results on SuperMicro-6012-p6, the article "2.6 kernel series
have noticeable performance regression, part 2" gives out the test
results on Toshiba tecra8000 laptop.

Re: 2.6 kernel series have noticeable performance regression

[Vincent Kergonna]

Hi,

Can we have the program/script you used to make your measures ?

Thanks.


On 6/14/05, yangyi <yang.yi@bmrtech.com> wrote:
> Hi, All
> 
> I did a performance regression test for all the 2.6 kernel series on two
> x86 platforms in last two weeks and found that some version kernels do
> have noticeable performance regression, some performance indices have
> about 20 to 30 percent regression,
>  even several indices have over 30 percent fall.
> 
> For example, on Toshiba tecra8000 laptop: for 4-bytes packet, network
> throughput of 2.6.11 is just about 57.5 percent of 2.6.0, for pipe
> bandwidth, the value of 2.6.11 is just about 68 percent of that value of
> 2.6.0, for mmap latency, the value of 2.6.11 is about 1.8 times as long
> as 2.6.0.
> 
> On supermicro-6012-p6, for 64-bytes small packet, network throughput of
> 2.6.11 is about 66.72 percent of 2.6.0, for 0K-size 2 processes context
> switch time, 2.6.11 is 31.8 percent bigger than 2.6.0, for pipe latency,
> 2.6.10 is 35.85 percent bigger than 2.6.3.
> 
> So, I think that it is very very necessary to find the problems and fix
> them ASAP.
> 
> The following articles give out all the test results, the article "2.6
> kernel series have noticeable performance regression, part 1" gives out
> the test results on SuperMicro-6012-p6, the article "2.6 kernel series
> have noticeable performance regression, part 2" gives out the test
> results on Toshiba tecra8000 laptop.
> 
> 
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
>

Re: 2.6 kernel series have noticeable performance regression

[yangyi]

LMbench
http://sourceforge.net/projects/lmbench

Netperf
ftp://ftp.cup.hp.com/dist/networking/benchmarks/netperf/

> Hi,
> 
> Can we have the program/script you used to make your measures ?
> 
> Thanks.
> 
> 
> On 6/14/05, yangyi <yang.yi@bmrtech.com> wrote:
> > Hi, All
> > 
> > I did a performance regression test for all the 2.6 kernel series on two
> > x86 platforms in last two weeks and found that some version kernels do
> > have noticeable performance regression, some performance indices have
> > about 20 to 30 percent regression,
> >  even several indices have over 30 percent fall.
> > 
> > For example, on Toshiba tecra8000 laptop: for 4-bytes packet, network
> > throughput of 2.6.11 is just about 57.5 percent of 2.6.0, for pipe
> > bandwidth, the value of 2.6.11 is just about 68 percent of that value of
> > 2.6.0, for mmap latency, the value of 2.6.11 is about 1.8 times as long
> > as 2.6.0.
> > 
> > On supermicro-6012-p6, for 64-bytes small packet, network throughput of
> > 2.6.11 is about 66.72 percent of 2.6.0, for 0K-size 2 processes context
> > switch time, 2.6.11 is 31.8 percent bigger than 2.6.0, for pipe latency,
> > 2.6.10 is 35.85 percent bigger than 2.6.3.
> > 
> > So, I think that it is very very necessary to find the problems and fix
> > them ASAP.
> > 
> > The following articles give out all the test results, the article "2.6
> > kernel series have noticeable performance regression, part 1" gives out
> > the test results on SuperMicro-6012-p6, the article "2.6 kernel series
> > have noticeable performance regression, part 2" gives out the test
> > results on Toshiba tecra8000 laptop.
> > 
> > 
> > -
> > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> > the body of a message to majordomo@vger.kernel.org
> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> > Please read the FAQ at  http://www.tux.org/lkml/
> >
>