Re: Network latency regressions from 2.6.22 to 2.6.29 (results with IRQ affinity)

[Eric Dumazet <dada1@cosmosbay.com>]

Eric Dumazet a écrit :
> Christoph Lameter a écrit :
>> On Thu, 16 Apr 2009, David Miller wrote:
>>
>>> We really need to find a good way to fix that IRQ affinity issue.
>> Here are the results with setting interrupt affinity. We are still loosing
>> 5 usecs from .22 to .30-rc2. We will investigate more in depth time
>> permitting.
>>
>> 	UDP ping pong 40 bytes
>>
>> Kernel		Test 1	Test 2	Test 3	Test 4	Avg
>> 2.6.22		83.1125	83.4	83.03	82.95	83.07
>> 2.6.23		83.33	83.51	83.18	83.42	83.21
>> 2.6.24		82.7775	82.8	82.77	82.77	82.77
>> 2.6.25		85.875	85.94	85.75	86.06	85.75
>> 2.6.26		87.4025	87.53	87.25	87.3	87.53
>> 2.6.27		87.81	87.81	87.77	87.82	87.84
>> 2.6.28		87.4275	87.77	87.24	87.28	87.42
>> 2.6.29		88.4325	88.48	88.42	88.44	88.39
>> 2.6.30-rc2	88.4925	88.49	88.44	88.56	88.48
> 
> 
> You are using UDP, and UDP got memory accounting. This 

This could explain a slowdown, but not 2.5 us per packet
(I take half your 5us, since you are measuring the two ways latencies)