We have often seen large delays in client perceived connection establishment latency, and don't have enough data to pin-down exactly where the delay is. While we often suspect delays may be happening in the TCP stack below mongoDB, or perhaps the network, we don't know how long MongoDB itself is taking on average to accept a new connection. 

As a first step should add a histogram that reveals how much time it takes for connections to be accept()ed on a socket by MongoDB's listener thread, until the connection is given its own dedicated thread and begins to run operations.