While running YCSB tests with multiple mongos instances on different machines I noticed that mongos seems to be opening double the number of connections to shards than client is opening to mongos.

So for example, client starts 75 threads, opens 75 connections to mongos and mongos ends up opening about 150 connections to each shard. (there are three shards and the workload are point updates and queries which are going to be evenly distributed across the shards).