There was originally a bottleneck reported on the CursorManager's mutex, tracked in SERVER-21754. However, after that was resolved we just moved the problem elsewhere, and there are reports of contention on the mutex for top, the session cache lock in wired_tiger_session_cache.h, and the mutex in src/mongo/db/s/sharding_state.h. This is nicely summarized in this comment on SERVER-21754.

As we saw with the cursor manager mutex, resolving the contention on those mutexes may also just push the problem to different mutexes, so this ticket is meant to track work to improve our read throughput when there are many cores (think 64+). These mutexes have been shown to be bottlenecks and are a good place to start, but there may be more work to do once those sources of contention are resolved.