We are using the 2.2.0-SNAPSHOT which from my understanding shuffles DNS
records across virtual users.
More precisely: the default is that each virtual user get a shuffled
version of what the JDK cache has.
There are a number of application servers but the CPU usage is highly
uneven. This could be a load balancer issue, it could be Gatling, it could
be something else.
Is this something you experience after elastically growing a cluster?
What is the best way of determining what requests go where?
FrontLine 
A packet analyser using grep and count?
At low load, you should see the IP in the debug log.