Re: [PATCH net-next 2/2] udp: implement and use per cpu rx skbs cache

From: Paolo Abeni <pabeni@redhat.com>
To: Eric Dumazet <eric.dumazet@gmail.com>, netdev@vger.kernel.org
Cc: "David S. Miller" <davem@davemloft.net>
Subject: Re: [PATCH net-next 2/2] udp: implement and use per cpu rx skbs cache
Date: Wed, 18 Apr 2018 19:15:12 +0200	[thread overview]
Message-ID: <1524071712.2599.60.camel@redhat.com> (raw)
In-Reply-To: <890db004-4dfe-7f77-61ee-1ac0d7d2a24c@gmail.com>

Hi,

On Wed, 2018-04-18 at 09:56 -0700, Eric Dumazet wrote:
> 
> On 04/18/2018 03:22 AM, Paolo Abeni wrote:
> > This changeset extends the idea behind commit c8c8b127091b ("udp:
> > under rx pressure, try to condense skbs"), trading more BH cpu
> > time and memory bandwidth to decrease the load on the user space
> > receiver.
> > 
> > At boot time we allocate a limited amount of skbs with small
> > data buffer, storing them in per cpu arrays. Such skbs are never
> > freed.
> > 
> > At run time, under rx pressure, the BH tries to copy the current
> > skb contents into the cache - if the current cache skb is available,
> > and the ingress skb is small enough and without any head states.
> > 
> > When using the cache skb, the ingress skb is dropped by the BH
> > - while still hot on cache - and the cache skb is inserted into
> > the rx queue, after increasing its usage count. Also, the cache
> > array index is moved to the next entry.
> > 
> > The receive side is unmodified: in udp_rcvmsg() the usage skb
> > usage count is decreased and the skb is _not_ freed - since the
> > cache keeps usage > 0. Since skb->usage is hot in the cache of the
> > receiver at consume time - the receiver has just read skb->data,
> > which lies in the same cacheline - the whole skb_consume_udp() becomes
> > really cheap.
> > 
> > UDP receive performances under flood improve as follow:
> > 
> > NR RX queues	Kpps	Kpps	Delta (%)
> > 		Before	After
> > 
> > 1		2252	2305	2
> > 2		2151	2569	19
> > 4		2033	2396	17
> > 8		1969	2329	18
> > 
> > Overall performances of knotd DNS server under real traffic flood
> > improves as follow:
> > 
> > 		Kpps	Kpps	Delta (%)
> > 		Before	After
> > 
> > 		3777	3981	5
> 
> 
> It might be time for knotd DNS server to finally use SO_REUSEPORT instead of
> adding this bloat to the kernel ?
> 
> Sorry, 5% improvement while you easily can get 300% improvement with no kernel change
> is not appealing to me :/

Thank you for the feedback.
Sorry for not being clear about it, but knotd is using SO_REUSEPORT and
the above tests are leveraging it.

That 5% is on top of that 300%.

Cheers,

Paolo