After I downloaded the sources of the JavaWS native stack 3.2.0.GA and debugged through the code, I was able to identify the piece of code that consumes so much time:
Class: ServiceDelegateImpl.java
Method: private <T> T getPortInternal(EndpointMetaData epMetaData, Class<T> seiClass)
Line 268: synchronized (epMetaData)
With one partner system it takes 1 to 2 seconds, with the other nearly 8 to 10 seconds.