Besides communication cost, what is the other source of inefficiency in RPC?

context switches, excessive buffer copying.
How can you optimise the communication?
(ans : communicate through shared memory on same machine, bypassing the kernel