This regression happened on "lkp-hsw-ep2", but “lkp-hsw-ep2” has been 
removed from the LKP tbox, so we retest the regression on another tbox 
“lkp-hsw-ep4”, their hardware is almost the same, "72 threads Intel(R) 
Xeon(R) CPU E5-2699 v3 @ 2.30GHz with 256G memory". The test result is 
in the attached file compare.txt. The regression is -6.7% for v5.6, the 
origin regression is -8.9%, it is not serious as it test on 
"lkp-hsw-ep2" before.


On 4/10/2020 6:14 PM, David Sterba wrote:
> On Fri, Apr 10, 2020 at 02:44:55PM +0800, Qu Wenruo wrote:
>> On 2020/4/10 下午2:34, Xing Zhengjun wrote:
>>> Hi Wenruo,
>>>
>>>     We test it in v5.6, the issue still exist, do you have time to take a
>>> look at this? Thanks.
>>
>> This is expected.
>>
>> The extra check brings new overhead mostly equal to another CRC32 run.
>>
>> We believe it's worthy, as our read time tree checker has exposed quite
>> some bit flip corruption.
> 
> The test probably runs on a PMEM device so there's no slowdown from the
> actual IO and the in-memory checks are measurable, though 28% is a lot,
> I'd expect something like 5-10% at most.
> 

-- 
Zhengjun Xing