Odd latency distribution as number of servers increases

I've wrote a micro benchmark for this at: https://github.com/Cjen1/prototyping-testing/tree/master/async_rpc

When increasing the number of server processes running I observed some slightly odd latency behavior:

I'm just wondering if the latency behavior is an artifact of my benchmark? Or is it a property of these kind of requests.

(Currently thinking that there is some fast path optimisation which is taken in some percentage of the requests to low numbers of servers)

Start with the async_rpc microbenchmark linked in the issue and reproduce the latency distribution while increasing the number of server processes. Check whether the observed pattern comes from the benchmark or from Async RPC request behavior. Done means providing a reproducible explanation, with supporting measurements or a confirmed issue in the RPC implementation.

Written by the indexing model from the issue text.