Many times load balancing can change what the optimal solution is for a given deployment.
Unless careful consideration is given to what is trying to be accomplished then some erroneous conclusions can be arrived at...
For example, people take it as gospel that netTcp binding is superior for performance... That might not be the case depending on how you're using it. Pooling and affinity are great...unless you want to scale in a round robin fashion.
Before you choose a binding for a connector you should think of your goals and then read the guidance here http://msdn.microsoft.com/en-us/library/ms730128.aspx.
After that you should test in your environment.
Sometimes, affinity is absolutely essential for lowest latency. In that case you may need a hybrid approach that allows for a dedicated connection via the API which dispatches according to need.
If your standard is to use web services for everything and not the API rememeber that you can use custom bindings and custom config files with Neuron so you should be able to tune for your needs.
I've attached a test program written by Jesse Ezell when he was trying to illustrate a point about our "Allow connection re-use" feature in our Service Connectors. Running it may surprise you....After you run it ask yourself what your use case is and what you're trying to accomplish. I think if you combine that with the information above and consider your solution's particular needs you'll be on the right track. Program.cs