We have an application that is clustered and cloned. When I say that, I mean it is clustered across two servers and it is cloned on those servers, so essentially I have 4 copies of the application running in their own JVMs. Every once in a while one of the JVMs will exceed its heap size and when it does it can prevent login that affects the other 3 JVMs. Sometimes it does what I call falls over and the dmgr restarts it but sometimes it just hangs and this can prevent access to the other 3 instances. Once you kill the pid (stopping is not always an option) then it frees up the other 3 instances immediately and life is good and you can login again. HA (High Availability) is supposed to hide these things from the user population but it doesn’t always. We are running WAS 8.5.5 ND.

In your description you have 4 JVMS running at the same time of which two are clones. I suspect the clones to have an influence in this scenario. You may want to experiment without the clones, fine tune the environment and then introduce the clones one at a time.. 

Alternatively you may want to locate each JVM on its on Virtual machine to ensure independence of the PID files.

Kindly correct me if am wrong.