What's the Fastest roundtrip communication using GM ?

Hi, everyone!
  I've some questions about the roundtrip performance of GM between nodes. 
For general messages, GM uses gm_send_with_callback/gm_directed_send_with
_callback, it involves about 9us latency as I tested. 
  But as I learn more about GM, I find that the mapper config or scout messages transmission
doesn't use these functions. They must be much more efficient for small messages communication between nodes.
  We are interested in using small messages(just a few bytes) to communicate between
the NIs just to find the status or load of the system. Since MCP can present some of the
information, the coming messages are not necessary to be delived to user space. So it
requires extreamly low latency. But the gm_send_with_callback/gm_directed_send_with fail
to do so.
  Any explainalation about the mapper config/scout messages communication or any advice about
the efficient communication of small messages between NIs are greatly appreciated.

Best regards
Geoge Zhang