The only chatting going on in a cluster is by
- FD (are-you-alive messages to one's neighbor node)
- STABLE (distributed garbage collection)
- MERGE(2): merge layer
You can reduce those by increasing the associated timeouts, in the case of FD you can repace it with FD_SOCK.
Other than that I suggest you set a WARN for org.jgroups, and only enable other cats at the INFO/DEBUG or even TRACE level when needed (debugging)