It seems that we currently do not have tests that also monitor the memory usage of mongos instances for sharded cluster workloads. Current tests seem to measure primarily the query latencies and throughputs.

There is a ticket for tracking and exposing the memory usage of AsyncResultsMerger instances in mongos (SERVER-105981). Once that is implemented, it would be reasonable to extend our existing performance tests for sharded queries so that they also track and expose the mongos usage. This will help us to find regressions in mongos memory usage in the future.