VLAN egress performance

VLAN egress performance

[Mattias Rönnblom]

Hi.

I'm running Linux on PC w/ a Core i7 CPU and two Intel 82598 NICs, and
I see some anomalies when it comes to egress VLAN performance. I
thought maybe someone on this list was interested in my results.

I'm running the stock Ubuntu 2.6.31 kernel, but with a newer ixgbe
driver (2.0.44.14).

The benchmark is IP forwarding with unidirectional UDP flows @ 64 byte
packets, and I get:

Ingress VLAN  Egress VLAN  Packet Rate    CPU utilization (all cores)
No            No           5.0 Mpacket/s  ~70%
Yes           No           5.0 Mpacket/s  ~75%
No            Yes          1.4 Mpacket/s  ~26%
Yes           Yes          1.3 Mpacket/s  ~26%

"VLAN" here mean I've put a VLAN device on top of the real ixgbe
device.

As you can see, if the egress i/f is a VLAN i/f, the performance is
reduced to less than a third. And in the case of egress VLAN, the
systems basically only uses one HW thread (with a softirqd process
taking up all the time).

Enabling lockdep, it looks like execution is serialized to a large
extent by contention around the "vlan_netdev_xmit_lock_key" lock.

I call this anomaly because I was surprised to see it (as oppose to
other performance degradations/scalability issues in the area of
multicore and IP traffic handling performance).

Does anyone know how difficult it would be to resolve this?  There's
not actually any synchronization needed between the different cores in
the case of VLAN. Correct? This is just an artefact of how the
implementation is done?

I looked briefly at the code, and I must admit I had some trouble
understanding where the flow in terms of locking.

Best regards,
     Mattias

Re: VLAN egress performance

[Patrick McHardy]

Mattias Rönnblom wrote:
> Hi.
> 
> I'm running Linux on PC w/ a Core i7 CPU and two Intel 82598 NICs, and
> I see some anomalies when it comes to egress VLAN performance. I
> thought maybe someone on this list was interested in my results.
> 
> I'm running the stock Ubuntu 2.6.31 kernel, but with a newer ixgbe
> driver (2.0.44.14).
> 
> The benchmark is IP forwarding with unidirectional UDP flows @ 64 byte
> packets, and I get:
> 
> Ingress VLAN  Egress VLAN  Packet Rate    CPU utilization (all cores)
> No            No           5.0 Mpacket/s  ~70%
> Yes           No           5.0 Mpacket/s  ~75%
> No            Yes          1.4 Mpacket/s  ~26%
> Yes           Yes          1.3 Mpacket/s  ~26%
> 
> "VLAN" here mean I've put a VLAN device on top of the real ixgbe
> device.
> 
> As you can see, if the egress i/f is a VLAN i/f, the performance is
> reduced to less than a third. And in the case of egress VLAN, the
> systems basically only uses one HW thread (with a softirqd process
> taking up all the time).
> 
> Enabling lockdep, it looks like execution is serialized to a large
> extent by contention around the "vlan_netdev_xmit_lock_key" lock.
> 
> I call this anomaly because I was surprised to see it (as oppose to
> other performance degradations/scalability issues in the area of
> multicore and IP traffic handling performance).

2.6.32 contains VLAN multiqueue support and should scale better.

Re: VLAN egress performance

[Eric Dumazet]

Le lundi 08 février 2010 à 14:13 +0100, Patrick McHardy a écrit :
> Mattias Rönnblom wrote:
> > Hi.
> > 
> > I'm running Linux on PC w/ a Core i7 CPU and two Intel 82598 NICs, and
> > I see some anomalies when it comes to egress VLAN performance. I
> > thought maybe someone on this list was interested in my results.
> > 
> > I'm running the stock Ubuntu 2.6.31 kernel, but with a newer ixgbe
> > driver (2.0.44.14).
> > 
> > The benchmark is IP forwarding with unidirectional UDP flows @ 64 byte
> > packets, and I get:
> > 
> > Ingress VLAN  Egress VLAN  Packet Rate    CPU utilization (all cores)
> > No            No           5.0 Mpacket/s  ~70%
> > Yes           No           5.0 Mpacket/s  ~75%
> > No            Yes          1.4 Mpacket/s  ~26%
> > Yes           Yes          1.3 Mpacket/s  ~26%
> > 
> > "VLAN" here mean I've put a VLAN device on top of the real ixgbe
> > device.
> > 
> > As you can see, if the egress i/f is a VLAN i/f, the performance is
> > reduced to less than a third. And in the case of egress VLAN, the
> > systems basically only uses one HW thread (with a softirqd process
> > taking up all the time).
> > 
> > Enabling lockdep, it looks like execution is serialized to a large
> > extent by contention around the "vlan_netdev_xmit_lock_key" lock.
> > 
> > I call this anomaly because I was surprised to see it (as oppose to
> > other performance degradations/scalability issues in the area of
> > multicore and IP traffic handling performance).
> 
> 2.6.32 contains VLAN multiqueue support and should scale better.

Yes, patches were added in September 2009, 2.6.32 should be OK.

commit 2f8bc32b7a08502a79e0ccec8697000f2977f2fd
Author: Eric Dumazet <eric.dumazet@gmail.com>
Date:   Thu Sep 3 02:19:58 2009 -0700

    vlan: enable multiqueue xmits
    
    vlan_dev_hard_start_xmit() & vlan_dev_hwaccel_hard_start_xmit()
    select txqueue number 0, instead of using index provided by
    skb_get_queue_mapping().
    
    This is not correct after commit 2e59af3dcbdf11635c03f
    [vlan: multiqueue vlan device] because
    txq->tx_packets  & txq->tx_bytes changes are performed on
    a single location, and not the right locking.
    
    Fix is to take the appropriate struct netdev_queue pointer
    
    Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
    Signed-off-by: David S. Miller <davem@davemloft.net>

commit 2e59af3dcbdf11635c03f22bfc9706744465d589
Author: Eric Dumazet <eric.dumazet@gmail.com>
Date:   Wed Sep 2 18:03:00 2009 -0700

    vlan: multiqueue vlan device
    
    vlan devices are currently not multi-queue capable.
    
    We can do that with a new rtnl_link_ops method,
    get_tx_queues(), called from rtnl_create_link()
    
    This new method gets num_tx_queues/real_num_tx_queues
    from real device.
    
    register_vlan_device() is also handled.
    
    Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
    Signed-off-by: David S. Miller <davem@davemloft.net>

Re: VLAN egress performance

[Mattias Rönnblom]

Patrick McHardy <kaber@trash.net> writes:

>> Ingress VLAN  Egress VLAN  Packet Rate    CPU utilization (all cores)
>> No            No           5.0 Mpacket/s  ~70%
>> Yes           No           5.0 Mpacket/s  ~75%
>> No            Yes          1.4 Mpacket/s  ~26%
>> Yes           Yes          1.3 Mpacket/s  ~26%
>> 
> 2.6.32 contains VLAN multiqueue support and should scale better.

OK, excellent. I tried out 2.6.32.7 and it did 4.0 Mpacket/s with VLAN
on both ingress and egress, and ~4.7 Mpacket/s without any VLANs.

Thanks,
     Mattias