Re: r8169 (+others ?) and note_interrupt performance hit on 2.6.30.x

[Michal Soltys <soltys@ziu.info>]

Jarek Poplawski wrote:
> Michal Soltys wrote, On 07/26/2009 08:43 PM:
> 
> ...
>> Anwyay, not sure if this is useful, so if there is anything better
>> I could do, just push me in the right direction.
> 
> Short test of 2.6.29? (Unless you prefer a bisection...)
> 
> Jarek P.
> --
> To unsubscribe from this list: send the line "unsubscribe netdev" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> 

I did first bisection, and the whole cpu load increases seem a bit more 
complicated than I originally thought.

I noticed at least 2 (possibly 3) moments where it increased (from 
original full 1 gbit at 20% - 50% load of one core)

(1) jump from 20% - 50% to 80% - 90%

(2) commit f11a377b3f4e897d11f0e8d1fc688667e2f19708
     r8169: avoid losing MSI interrupts

After the above commit, one core is not enough to handle 1 gbit speed 
with 1k packets, it ends with with ~760mbit and one core drained at 100% 
. The situation right before this commit is as in (1)

All tests were done, as mentioned earlier - with nc process and eth 
interrupts pinned to the same core, and besides nc process - idle 
machine. Tested with udp only, standard pfifo_fast.

The changes happened between 2.6.29 and 2.6.30-rc6 .

I'll try to pinpoint when (1) happens now. Lookint at mpstat output, 
those may actually be two different commits.