Re: [rcu] c0f489d2c6f: -1.5% netperf.Throughput_tps

[Aaron Lu <aaron.lu@intel.com>]

[-- Attachment #1: Type: text/plain, Size: 1770 bytes --]

On 07/25/2014 05:44 PM, Mike Galbraith wrote:
> On Fri, 2014-07-25 at 16:05 +0800, Aaron Lu wrote: 
>> On 07/25/2014 03:35 PM, Mike Galbraith wrote:
>>> On Fri, 2014-07-25 at 14:45 +0800, Aaron Lu wrote: 
>>>> FYI, we noticed the below changes on
>>>>
>>>> git://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git master
>>>> commit c0f489d2c6fec8994c642c2ec925eb858727dc7b ("rcu: Bind grace-period kthreads to non-NO_HZ_FULL CPUs")
>>>>
>>>> abaa93d9e1de2c2  c0f489d2c6fec8994c642c2ec  
>>>> ---------------  -------------------------  
>>>>      12654 ~ 0%      -1.5%      12470 ~ 0%  ivb43/netperf/300s-25%-TCP_CRR
>>>>      12654 ~ 0%      -1.5%      12470 ~ 0%  TOTAL netperf.Throughput_tps
>>>
>>> Out of curiosity, what parameters do you use for this test?  In my
>>
>> The cmdline for this test is:
>> netperf -t TCP_CRR -c -C -l 300
> 
> Thanks.  That doesn't switch as heftily as plain TCP_RR, but I'd still
> expect memory layout etc to make bisection frustrating as heck.  But no
> matter, I was just curious.

The bisect is done by the LKP test system(developed by Fengguang)
automatically so it's not that painful for me :-) But as you have said,
the 1.5% change is too small and probably doesn't worth a report, I'll
be more careful next time when examining the report.

> 
> Aside: running unbound, the load may get beaten up pretty bad by nohz if
> it's enabled.  Maybe for testing the network stack it'd be better to
> remove that variable?  Dunno, just a thought.  I only mention it because

The CONFIG_NO_HZ_FULL is set to y, I'll disable it to see if the number
changes, thanks for the tips.

Regards,
Aaron

> your numbers look very low unless the box is ancient or CPU is dinky.
> 
> -Mike
>