Update: Industry db benchmark result

Update: Industry db benchmark result

[Chen, Kenneth W]

It's been awhile since last time I reported Linux kernel performance
for an industry standard transaction processing database benchmark. It's
time to report the data again to keep everyone informed.  Here is the
latest result that we measured (4P 1.6GHz ia64, 9M L3 cache, 64GB memory,
480 15K-RPM disks).

2.4.21-4.EL	baseline
2.6.9		- 6.00%
2.6.11	-13.00%
2.6.12-rc3	-13.54%
2.6.12-rc4	-13.05%
2.6.12-rc5	-12.86%
2.6.12-rc6	-12.66%

It's been hovering around -13%.  I will take a measurement of 2.6.12, though
I don't expect it to deviate too far from 2.6.12-rc6.  There are 718 patches
committed since 2.6.12 as of 6/22 14:09 PDT, so I'm going to take a snapshot
today or tomorrow and queue it up for benchmark measurement.  I will probably
take all the scheduler patches currently in -mm and give that a whirl as well.

- Ken

RE: Update: Industry db benchmark result

[Chen, Kenneth W]

Andi Kleen wrote on Wednesday, June 22, 2005 4:44 PM
> "Chen, Kenneth W" <kenneth.w.chen@intel.com> writes:
> > 
> > It's been hovering around -13%.  
> 
> Do you have an educated guess where the 13% loss is comming from? 

Process scheduler is the prime suspect. Earlier scheduler experiments
on top of 2.6.11 showed that it was about 10% regression.  Similar
experiments repeated on top of 2.6.12-rc4 indicated about a much
smaller 4%.  Other area are O_DIRECT (2%), and possibly scsi.

- Ken

Re: Update: Industry db benchmark result

[Andi Kleen]

"Chen, Kenneth W" <kenneth.w.chen@intel.com> writes:
> 
> It's been hovering around -13%.  

Do you have an educated guess where the 13% loss is comming from? 

-Andi