Re: [RFC PATCH 00/12] drop the qdisc lock for pfifo_fast/mq

[John Fastabend <john.fastabend@gmail.com>]

On 16-01-06 05:14 AM, Jamal Hadi Salim wrote:
> 
> Sorry for not being as responsive as i would like to be
> (theman calls and i have to go).
> This looks like a good (tc workshop) candidate discussion, if still
> active by netdev11 time.
> 
> On 15-12-30 12:50 PM, John Fastabend wrote:
>> Hi,
>>
>> This is a first take at removing the qdisc lock on the xmit path
>> where qdiscs actually have queues of skbs. The ingress qdisc
>> which is already lockless was "easy" at least in the sense that
>> we did not need any lock-free data structures to hold skbs.
>>
> 
> I did some testing over the holidays for a netdev11 paper submission
> and the ingress qdisc side of things looks very impressive (on
> an average packet size) on a single (i7 class) cpu.
> It can handle about 3x what an egress side pktgen type test (not
> very real life) can handle. Analysis shows the lock is killing us.
> So if you are looking at low hanging fruit, the egress is
> the place to look.

So this series drops the qdisc lock which should hopefully per
some older data I have bring it on par with the ingress side. But
I need to retest this latest patch series.

> I have a pktgen change that may be useful for you - I will post
> it next time i get cycles.
> I am also a willing guinea pig (given upcoming netdev11) to do
> some perf testing ;->
> 

I typically test this by putting vlan's on top of a dev with the
qdisc I want to test and running pktgen on the vlans the entry point
of the skb is pretty much where I want it in this case.

> cheers,
> jamal
> 
>