[ovs-dev] [PATCH] bond: Send gratuitous ARPs when load balancing.

Sanjay Sane ssane at nicira.com
Sun Jul 3 18:22:00 UTC 2011


Ben: what happens when link failover happens as a result of link carrier
going down ?

The 2 issues which are related are
1. duplicates to the host
2. packet blackholing

for issue #1

L2 network very frequently un-learns all the MAC addresses, this happens
whenever there is a topology change. We recently observed (in one customer
case), that L2 switch learning was not even functioning well. In either of
these cases, we'll get duplicates towards the host/VM mac address. This is
something we must ensure we dont pass on to the host.

in active-backup mode, we're dropping all packets we received on the backup
link.
in SLB mode, for a given MAC address, the scenario looks similar to the
active-backup mode. i.e for a given MAC, only one link should be active, and
I should drop all packets received on the backup link.

for issue #2

if we "enforce" no duplicates, by dropping packets on the backup link, we
must ensure that the L2 network at least sends the packet (destined to the
host MAC) on the correct link. The decision to change the "active" link
should therefore be communicated to the L2 network. if this decision was due
to the outgoing traffic, then obviously there's no need to do anything more.
If this decision was due to link failure related re-balancing, then we
should make use of gratuitous ARP, or learning packets.


-------
in layer-2 networks, duplicates are to be avoided, as traditionally it has
been an indication of some loops in the network. Packet blackholing must
also be avoided, of course, but given the choice of {loops, duplicates} Vs
"longer backholing", L2 networks typically have made a choice to avoid loops
or duplicates.

------

thanks
Sanjay


On Fri, Jul 1, 2011 at 2:08 PM, Ben Pfaff <blp at nicira.com> wrote:

> On Fri, Jul 01, 2011 at 02:04:38PM -0700, Jesse Gross wrote:
> > On Fri, Jul 1, 2011 at 11:07 AM, Ethan Jackson <ethan at nicira.com> wrote:
> > > I'm not particularly convinced that this is important, but Sanjay
> > > thinks it is. ??It was easy enough to write up so I thought we could
> > > discuss it on the dev list. ??I'm not married to the patch either way.
> > > ??Though agree I don't really see the value in forcing traffic to
> > > ingress the same link we send it on.
> >
> > Well the value is load balancing of incoming traffic.  Convincing the
> > remote switch to send traffic on the same link as is used for outgoing
> > packets is the only way that we have to do load balancing for SLB.
> >
> > This is really only beneficial for largely unidirectional streams
> > where we're not outputting traffic that would cause the remote switch
> > to figure out the correct link by itself.
>
> The SLB rebalancing logic is based solely on outgoing traffic, and we
> only move MACs that have a significant amount of outgoing traffic, so
> sending a gratuitous ARP is normally redundant.  The switch will find
> out that the MAC has moved on the next packet we send, which will
> normally be quite soon (since the MAC that was moved has a significant
> amount of outgoing traffic).
> _______________________________________________
> dev mailing list
> dev at openvswitch.org
> http://openvswitch.org/mailman/listinfo/dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.openvswitch.org/pipermail/ovs-dev/attachments/20110703/0c35a2d4/attachment-0003.html>


More information about the dev mailing list