Re: RSS is not efficient when forwarding (ixgbe)

From: Flavio Leitner <fbl@redhat.com>
To: Tom Herbert <therbert@google.com>
Cc: Linux Netdev List <netdev@vger.kernel.org>,
	Jacob Keller <jacob.e.keller@intel.com>,
	Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Subject: Re: RSS is not efficient when forwarding (ixgbe)
Date: Mon, 7 Jul 2014 07:55:23 -0300	[thread overview]
Message-ID: <20140707105523.GC2343@t520.home> (raw)
In-Reply-To: <CA+mtBx-dY2eoDzL-1K8F8Br224H+c4fzVtVPu9JmnjLdQKuAOg@mail.gmail.com>

On Sat, Jul 05, 2014 at 09:09:34AM -0700, Tom Herbert wrote:
> On Fri, Jul 4, 2014 at 12:36 PM, Flavio Leitner <fbl@redhat.com> wrote:
> > On Thu, Jul 03, 2014 at 07:44:08PM -0300, Flavio Leitner wrote:
> >>
> >> Hi,
> >>
> >> I have a simple router setup which forwards traffic from
> >> one ixgbe 82599ES to another ixgbe of the same model.
> >>
> >> kernel: 3.16.0-rc2-00262-ga921e2a
> >>
> >>   p2p1: 192.168.155.1/24 is the gateway of the LAN
> >>   p2p2: 192.168.156.1/24 is the gateway of the other LAN
> >>
> >> While the ARP is resolving, I can see the packets being spread
> >> among all the 8 queues (8 online CPUs) available and that is fine.
> >>
> >> However, as soon as the TCP traffic starts, all streams are
> >> merged to rx-queue-0 which overwhelms one single CPU, so the
> >> total throughput is about 4Gbits/sec.
> >>
> >> I can see the driver sending different skb->hash for each stream,
> >> so it can't be the NIC.
> >>
> >> Also, if I run a local http on the router, the skb->hash pattern
> >> doesn't change, but the workload is spread among all CPUs.
> >>
> >> debug output while reproducing all the streams on rx-queue-0:
> >> [...]
> >> [11685.885093] ixgbe_rx_skb:1713 skb(ffff880222a77200) hash: 0xC2AF4A27
> >> [11685.891454] ixgbe_rx_skb:1713 skb(ffff880222a77200) hash: 0x8C5B749D
> >> [11685.897820] ixgbe_rx_skb:1713 skb(ffff880222a77200) hash: 0xA33BA6D5
> >> [11690.845032] net_ratelimit: 3276406 callbacks suppressed
> >> [...]
> >
> > The command 'ehtool -S p2p1 | grep rx_queue' shows only rx queue #0
> > receiving packets.  Also, /proc/interrupts shows only p2p1-TxRx-0
> > generating interrupts.
> >
> > Nothing changes when I start irqbalance in the middle of the test.
> > However, if I create new streams, they are correctly distributed
> > among the NIC queues/irqs/CPUs.
> >
> Could this be interaction with flow director? Is the application
> running on just one CPU?

It's a router forwarding traffic from one interface to another, so
I guess it's only the kernel. BTW, no firewall.

Flow Director needs to be enabled and I am using defaults.

> > I have tried the same setup and test with another card (bnx2, 1GbE)
> > and all queues got traffic by default though they were all assigned
> > to CPU#0 (no irqbalance), so this is expected and makes me think
> > this is specific to the ixgbe driver.
> >
> > Any ideas?
> >
> > Thanks,
> > fbl
> > --
> > To unsubscribe from this list: send the line "unsubscribe netdev" in
> > the body of a message to majordomo@vger.kernel.org
> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> --
> To unsubscribe from this list: send the line "unsubscribe netdev" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> 