[ovs-discuss] Bonding problems with redundant physical switching

Ben Pfaff blp at nicira.com
Tue Sep 14 21:59:38 UTC 2010


On Fri, Sep 10, 2010 at 10:48:12PM +0200, Christian Fischer wrote:
> On Friday September 10 2010 21:36:50 Ben Pfaff wrote:
> > On Fri, Sep 10, 2010 at 09:23:30PM +0200, Christian Fischer wrote:
> > > I try to get nic bonds working with OpenvSwitch-0.99.2 (XCP-0.5.0).
> > > Due to missing xcp administration documentation I've started with the
> > > XenServer-5.6 administration guide.
> > >
> > > There is stated that SLB bonding is an active/active mode.
> > >
> > > If I create bonds with OpenvSwitch then they are active/passive.
> >
> 
> Thanks Ben for the reply.
> 
> > The SLB bonding implemented by Open vSwitch is indeed active/active.
> > Please explain why you think that it is active/passive.
> 
> The bond is not reponding to packages arriving eth1, that's why I thought eth1 
> is passive.
> 
> # ovs-appctl bond/show bond0
> updelay: 200 ms
> downdelay: 0 ms
> next rebalance: 562 ms
> slave eth1: enabled
> slave eth0: enabled
>         active slave
>         hash 44: 1 kB load
>                 00:15:17:de:dd:c8
> 
> Is it active/active for sending or for sending and receiving?

It is active/active for sending and receiving.  (Packets from any given
source MAC are assigned to a particular slave though.)

Broadcasts and multicasts are only accepted on the "active slave",
however; otherwise such packets would be received in duplicate.  And
there are a few special rules for similar corner cases, if I recall
correctly.

> > > I've redundant physical switching for the bond networks, but without
> > > spanning tree support (HP 1810G), therefore no inter switch links. If
> > > on one side NIC0 active and on the other side NIC1 active then
> > > connectivity is broken.  Normally i would think that it should be
> > > possible to create a trunk between two nodes running identical trunk
> > > mode without any switching, in this case there's also no ISL.
> >
> > I don't understand this paragraph.
> 
> All physical cluster nodes have different physical networks, 2 network 
> interfaces bonded each. There are two physical switches, the networks have 
> different VlanIDs. There's no network link between both switches, the 
> switches have no spanning tree protocol support.

The important question here is, are the two links within a bond
connected to the same switch?  They should be; that is the intended and
supported configuration.

> Shouldn't both slaves listen for incoming packages?

Yes; they do (except for broadcast and multicasts; see above).

> What should happen if one connects two hosts with two bonded NICs each 
> directly without any switches?

In other words, host0 has bond0 composed of eth0 and eth1, and host1 has
bond1 composed of eth2 and eth3, eth0 is connected to eth2 with a cable,
and eth1 is connected to eth3 with a cable?  I think that should work
OK.




More information about the discuss mailing list