You have a data center standardized on 10GE connections. With, say, Nexus 7000s in the core, Nexus 5000s at the aggregation, and some fabric extenders for the edge to the servers (I use Cisco gear as examples because this is what's in my specific lab). There are some ACE 4710 load balancers hanging off your Nexus 5000s, but these only have 1GE interfaces. All of your switch connections are 10GE, needed for the massive east-west (VM-to-VM) traffic in modern virtualized data centers.

Do the load balancers not become a bottleneck in certain traffic conditions? I can see how some local east-west traffic does not need to even reach the load balancer, but there are other situations where you need to traverse the core, and possibly even a data center interconnect.

Basically, I know the load balancers are used in client-server (north-south) traffic, and things like HTTP GET do not need 10GE, but are there situations in which your 1GE load balancer can get in the way of an otherwise all-10GE traffic path and cause problems for things like vMotion (for example)?

  • 1
    The ACE 4710 has been EOL/EOS since 2010. Are you tied to this load balancer, or are you open to using a modern load balancer that can scale much higher? (Many manufacturers make them.) There is no reason to artificially place limits on your throughput in the way you describe. Well, no reason outside of cost, but really if you've bought the infrastructure you describe, you would likely have cash left over for a real load balancer setup. – Brett Lykins Mar 27 '14 at 16:07
  • This is actually not a production environment, but an actual "lab" scenario that contains these particular devices. In a more general sense, my question is, given the traffic load of a typical data center, how do you ensure the load balancer does not become a bottleneck in your design? This plays in to your design, like for example if your data center is multitiered, at what tier do you do load balancing, etc. In my particular example, which I cannot change since it's not my design, does it make sense to have these ACE devices one-armed off the Nexus 5000s who are much more powerful. – sentinel Mar 27 '14 at 17:12

Do the load balancers not become a bottleneck in certain traffic conditions?

Certainly, but this is generally not the case in a well engineered network.

You should be able to design your network in a way that would allow the majority of your internal server-to-server traffic ("east-west" as you put it) even if it does need to cross your core or between data centers.

While often the load balancer is the default gateway for the servers behind it, I have seen setups where a routing protocol (i.e. OSPF or RIP) is run to allow the "east-west" traffic to circumvent the load balancer, or in smaller deployments where static routes were used.

If the load balancers are going to be a bottleneck even in with a good design (i.e. the volume of traffic is just that high), then there are ways of load balancing across multiple load balancers as well.

| improve this answer | |
  • 4
    Indeed. If there's a LB in your vMotion path, you've failed as an engineer. – Ricky Beam Mar 27 '14 at 19:28
  • "... there are ways of load balancing across multiple load balancers as well" - such as selecting the network load balancer from the DNS? – Andrei Rînea Mar 14 '17 at 12:09

This is indeed a bottleneck and will limit your throughput to the "weakest link in the chain" which is your LBs. However, you can get around it. If you use something known as "switchback" or "direct server return" you can do async traffic flows. The way it works is this:

a) client makes a http request to

b) LB answers on and passes the incoming request to a server -- since the LB and server are on the same LAN, this is done at layer 2.

c) The server accepts this incoming connection as the IP is on an aliased loopback interface (otherwise it'd drop a packet that didn't match an interface).

d) The server responds directly to the client.

The request is a few bytes. The content served can be of any size. The outgoing traffic doesn't go through the LB so you can handle WAY more traffic.



| improve this answer | |
  • 1
    Keep in mind that doing so is only appropriate for L4 clusters. For L7 clusters, this will break header rewrites, cookie insertion (persistence), SSL termination, any sort of URL matching, and any other more advanced features of many load balances. – YLearn Mar 30 '14 at 16:23

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.