This library is eye-catching. awesome! How high is the call overhead when used on a single machine? Would it be faster if communication is done through shared memory? How fast does a client cold start?