The replica set monitor in older versions of the server initializes it's random state via:
which has the unfortunate side effect of being strongly synchronized with the system clock. This can cause non-random patterns of access across a mongos / series of mongos' in cases where all of the nodes are restarted at the same time and work is distributed evenly across the various mongos'.
Switching to use a real source of randomness for seeding the prngs in each set state should improve the real randomness of later server selection
Note: It appears this was fixed in 4.2 as part of unrelated code changes