A couple of comments.
-Xmn2g: I would not let the young generation get up to half the heap size, you too often end up with too many full collection. I'd limit it to 1gb (I usually recommend 1/3 to 1/4 the heap size as a starting point for tuning). This is especially important when using the CMS collector because while the CMS collector is running you will loose the use of one of your cores because it will be busy running the collector.
-XX:ParallelGCThreads=8: I hope that you have at least 8 cores. Never set this to more cores than you have. The JVM by default sets this to the number of cores (for up to 8 cores, but after that it does 6 threads per 8 cores IIRC.)
You still haven't figured out the overheap introduced by Axis2 (not sure why you keep on ignoring this, especially since it is critical to resolving the issue you keep on brining up)