Your browser failed to load CSS style sheets. Your browser or web proxy might not support elliptic-curve TLS

Building network automation solutions

6 week online course

Start now!
back to overview

Don’t try to fake Multi-chassis Link Aggregation (MLAG)

Martin sent me an interesting challenge: he needs to connect an HP switch in a blade enclosure to a pair of Catalyst 3500G switches. His Catalysts are not stackable and he needs the full bandwidth between the switches, so he decided to fake the multi-chassis link aggregation functionality by configuring static LAG on the HP switch and disabling STP on it (the Catalysts have no idea they're talking to the same switch):

He was already aware of at least two problems (but, as a vendor product manager told me in a recent phone call, "that's nothing that couldn't be solved with a bit of scripting"):

  • If one of the links fails, static LAG will not detect link failure and some of the servers connected to the HP switch will lose connectivity to the network.
  • If the link between the Catalysts fails, he’ll face a split-brain scenario, as they won’t be able to use the connectivity through the HP switch.

In my opinion that should have been more than enough to run away from this “solution”, as I prefer to sleep during the night and like my network to survive single failures, but it only gets worse:

Forwarding loop following a configuration change. An overworked engineer trying to troubleshoot connectivity problems at 2AM might decide to disable LAG on the HP switch after an inter-switch link failure. The following morning, someone else (unaware of the configuration change) fixes the physical link and the whole network starts glowing in the dark.

Load balancing on the static LAG. If you configure source-MAC-based load balancing on the HP switch (X in my diagram), things just might work (of course we still won’t be able to survive a single link failure). However, if someone figures out a server is overloading its uplink and configures source-destination-MAC-based load balancing or (even worse) per-session load balancing on the LAG, the Catalysts (C1 and C2 in my diagram) will go crazy screaming about flapping MAC addresses (yet again, thanks to all of you who helped me understand this particular problem).

Suboptimal traffic flow. Assuming load balancing in X associates server B is with the X-C1 link, the traffic between D and B will flow over C2, C1 and X due to the way MAC address learning works.

Replicated multicasts. When a host attached to this network sends a broadcast (or multicast or unicast to unknown destination), all the hosts attached to X will receive two copies of the packet. For example, if A sends an ARP request (broadcast), C1 will propagate it to C2 and X and C2 will propagate a second copy to X. B and C will thus receive two copies of the same ARP request.

ARP will obviously work, as it has been designed for lossy unreliable networks. Some fainthearted protocols assuming Ethernet is lossless (I’m looking at you, FCoE) might not be so robust. Figuring out how various server clustering kludges (that rely on layer-2 connectivity between cluster nodes) work in this scenario would also be great fun.

Last but not least, if B sends a broadcast packet, it will receive a copy of its own broadcast propagated through X-C1-C2 (or X-C2-C1).

Anything else? Is there something else that can go wrong? Please let me know in the comments!

More information

Multi-chassis Link Aggregation (MLAG) technologies and data center architectures from Cisco, Juniper and HP are described in my Data Center 3.0 for Networking Engineers webinar (register for an online session or buy a recording).


  1. Ivan,
    Can you please clarify something for me?
    You say "If one of the links fails, static LAG will not detect link failure". Are you saying that if we don't use LACP or PAGP and instead use 'channel-group 1 mode on', for example, and a link in etherchannel fails, some traffic will be dropped? As in packets are still load balanced over the dead link??
    Or is this an HP thing?

  2. Ivan Pepelnjak21 March, 2011 20:00

    This is not an HP thing. If you have static LAG and a member link fails in non-obvious way (forwarding loss with no carrier loss), you'll get a traffic black hole. LACP has a keepalive mechanism and thus detects that one of the links in LAG has failed and removes it from LAG.

  3. NotsoAnonymous22 March, 2011 03:22

    The article should have ended where it said

    "configuring static LAG on the HP switch and disabling STP on it (the Catalysts have no idea they're talking to the same switch)"


  4. According to my experiences HP switches forward STP BPDUs if STP is disabled. Thus, the Cisco switches see each others BPDUs in two different ports and one of those is made STP blocking. If the blocking port is a port facing X, then ~half of the traffic sent to upstream by X is blackholed.

  5. (assuming source-MAC-based load balancing on X)

  6. Markku Leiniö22 March, 2011 21:34

    Oh, and in this particular topology, having the C1-C2 link blocking obviously blocks something else as well... (A and D will not necessarily see each other at all)

  7. Ivan Pepelnjak24 March, 2011 20:08

    It's not that bad. The HP switch will not send the multicast packet (BPDU) back across the same LAG, so the BPDUs sent by the Catalyst switches to the STP-ignoring HP switch will not arrive at the other Catalyst switch.

  8. Ivan Pepelnjak24 March, 2011 20:09

    And your point is ... what?

  9. You're right, my mistake. I apparently don't like too imaginative configs... and have been dealing with too many ugly ProCurve details lately.

  10. Ivan

    You forgot to mention Nortel (now Avaya) when you say Chassis Aggregation technologiees. Nortel does SMLT since 2001.


You don't have to log in to post a comment, but please do provide your real name/URL. Anonymous comments might get deleted.