Re: High contention on the sk_buff_head.lock

[Gregory Haskins]

David Miller wrote:
> From: Vernon Mauery <vernux@xxxxxxxxxx>
> Date: Wed, 18 Mar 2009 14:10:33 -0700
>
>   
>> David Miller wrote:
>>     
>>> From: Andi Kleen <andi@xxxxxxxxxxxxxx>
>>> Date: Wed, 18 Mar 2009 21:54:37 +0100
>>>
>>>       
>>>> But then again I'm not sure it's  worth it if the problem only
>>>> happens in out of tree RT.
>>>>         
>>> The list of problems that only show up with the RT kernel seems to be
>>> constantly increasing, but honestly is very surprising to me.
>>> I don't understand why we even need to be concerned about this stuff
>>> upstream, to be honest.
>>> Please reproduce this in the vanilla kernel, then get back to us.
>>>       
>> Huh? The numbers that I posted *were* from the vanilla kernel. I ran
>> the 2.6.29-rc8 kernel with lock_stat enabled.  The lock contention
>> happens on the same lock in both vanilla and -rt, it just happens
>> to be more pronounced in the -rt kernel because of the double context
>> switches that the sleeping spinlock/rt-mutexes introduce.
>>     
>
> And the double context switches are probably also why less
> natural batching and locality are achieved in the RT kernel.
>
> Isn't that true?
>   

Note that -rt doesnt typically context-switch under contention anymore
since we introduced adaptive-locks.  Also note that the contention
against the lock is still contention, regardless of whether you have -rt
or not.  Its just that the slow-path to handle the contended case for
-rt is more expensive than mainline.  However, once you have the
contention as stated, you have already lost.

We have observed the posters findings ourselves in both mainline and
-rt.  I.e. That lock doesnt scale very well once you have more than a
handful of cores.  It's certainly a great area to look at for improving
the overall stack, IMO, as I believe there is quite a bit of headroom
left to be recovered that is buried there.

Regards,
-Greg

Attachment: signature.asc
Description: OpenPGP digital signature