Re: [net-next PATCH 1/3] net: make default tx_queue_len configurable

[Phil Sutter <phil@nwl.cc>]

On Tue, Aug 11, 2015 at 06:13:49PM -0700, Alexei Starovoitov wrote:
> On Tue, Aug 11, 2015 at 06:23:35PM +0200, Phil Sutter wrote:
> > 
> > I have an unfinished solution in the oven, but being kept busy with
> > other things for now. The action plan is as follows:
> > 
> > 1) Introduce IFF_NO_QUEUE net_device->priv_flag.
> > 2) Have attach_default_qdiscs() and attach_one_default_qdisc() treat
> >    IFF_NO_QUEUE as alternative to tx_queue_len == 0.
> > 3) Add warning to register_netdevice() if tx_queue_len == 0.
> > 4) Change virtual NIC drivers to set IFF_NO_QUEUE and leave tx_queue_len
> >    alone.
> > 5) Eventually drop all special handling for tx_queue_len == 0.
> > 
> > I am currently somewhere in 2) and need to implement 4) for veth as PoC to
> > check if 2) suffices in all situations we want. Not sure if 3) is
> > desireable at all or if there are valid cases for a literally zero
> > length TX queue length.
> 
> sounds like you want to change default qdisc from pfifo_fast to noqueue
> for veth, right?
> In general 'changing the default' may be an acceptable thing, but then
> it needs to strongly justified. How much performance does it bring?
> Also why introduce the flag? Why not just add 'tx_queue_len = 0;' 
> to veth_setup() like the whole bunch of devices do?

A quick test on my local VM with veth and netperf (netserver and veth
peer in different netns) I see an increase of about 5% of throughput
when using noqueue instead of the default pfifo_fast.

Cheers, Phil