[f-nsp] ALB + VRRP

[George B.]

I have had all sorts of different issues doing this.  The sort of issues you
have will also depend on which bonding driver you are using (Ubuntu, for
example, sometimes uses an older bonding driver than the one that comes with
the stock kernel of a given version).  I am assuming you are using the Linux
bonding driver for this in something like the balance-alb mode (mode 6).
The problem with certain kernels seems to be that when a packet arrives at
the NIC that has a different MAC address than the bond (the MAC of the
secondary slave), it is ignored.    So it does what it is supposed to do on
the transmit side (overwrites the MAC address of the bond device ... which
is the primary slave ... with that of the secondary slave ... but any
replies to that MAC address get ignored).

I have given up on that method completely and use two stacked FCX switches
with the host in balance-xor mode.  There were just too many different
problems that varied from one kernel/driver version to the next and when you
add to the mix that the various distributions tend to fiddle with things, it
quickly becomes a mess and what might work with one version breaks the next
time a kernel package is released.



On Thu, Dec 9, 2010 at 12:02 PM, Nick Morrison <nick at nick.on.net> wrote:

>
> Hi,
>
> Has anyone got any first-hand experience with servers running ALB (adaptive
> load balancing is what Linux calls it - windows machines know it by other
> names, maybe "active active" load sharing - where the host is connected
> with multiple NICs to multiple edge switches (same segment), and
> round-robins its ARP replies so that different hosts on the same segment
> know the single IP by different MACs) ... in conjunction with RX routers
> running VRRP-E?
>
> I'd like to hear about your experiences.  I'm having a few issues that look
> like our two VRRP-E RX routers are trying to be smarter than I want them to
> be (ARP snooping, but buggering it up so that their ARP table ports don't
> match their MAC table ports, and so sending traffic destined to one of the
> NICs on the server out of the wrong port) but it's hard to pinpoint the
> cause of the problem...
>
> Topology is two edge RX-8s, cross-connected to two distribution RX-8s.
> Edge RX-8s just do L2, distro RX-8s hold the default gateway and do VRRP-E.
> Host is connected to both edge RX-8s.  One IP address, two MAC addresses
> (one per NIC), same VLAN, both NICs up and active, host manages ARP replies
> carefully.  (In theory.)
>
>
> Cheers,
>
> Nick!
>
> _______________________________________________
> foundry-nsp mailing list
> foundry-nsp at puck.nether.net
> http://puck.nether.net/mailman/listinfo/foundry-nsp
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://puck.nether.net/pipermail/foundry-nsp/attachments/20101209/e8432925/attachment.html>