Re: [PATCH] Re: Bad network performance over 2Gbps

["Kok, Auke" <auke-jan.h.kok@intel.com>]

Bodo Eggert wrote:
> On Mon, 21 Apr 2008, Chris Snook wrote:
>> Bodo Eggert wrote:
>>> On Mon, 21 Apr 2008, Rick Jones wrote:
>>>> Bodo Eggert wrote:
>>>>> On Mon, 21 Apr 2008, Rick Jones wrote:
> 
>>>>>> Be it kernel or user space, for consistent benchmark results it needs to
>>>>>> be
>>>>>> able to be turned-off without turning the code.  That leaves me in
>>>>>> agreement
>>>>>> with Stephen that if it must exist, the user space one would be
>>>>>> preferable.
>>>>>> It can be easily terminated with extreme prejudice.
>>>>> I agree that having a full-featured userspace balancer daemon with lots of
>>>>> intelligence will be theoretically better, but if you can have a simple
>>>>> daemon doing OK on many machines for less than the userspace daemon's
>>>>> kernel stack, why not?
>>>> Perhaps my judgement is too colored by benchmark(et)ing, and desires to have
>>>> repeatable results on things like neperf, but I very much like to know where
>>>> my interrupts are going and don't like them moving around. That is why I am
>>>> not particularly fond of either flavor of irq balancing.
>>>>
>>>> That being the case, whatever is out there aught to be able to be disabled on
>>>> a running system without having to roll bits or reboot.
>>> Adding a "module" parameter to disable it should be cheap, isn't it?
>> Except the irq balancing is system-wide.  Adding per-device exemptions to an 
>> obsolete feature seems like the wrong way to go.
> 
> No, not a per-device-exemption. My reasoning was: If the IRQ balancer 
> bounces the IRQ too often, doing it less often seems to be the correct 
> solution. One cache miss each ten seconds sounds like it should be OK.
> As said before, I can't verify this theory.

this is exaclty what the userspace irqbalance does and it's even optimized to not
do those migrations once every 10 seconds if things look OK. from that
perspective, it's definately more mature and it's maintained as well.

Auke