1-node repl set, HDD, 500 GB oplog, 4 GB cache, 5 threads inserting 130 kB docs. Left is 3.6.0-rc8, right is 3.4.10:
At the point A where the oplog fills and oplog truncation begins in 3.6.0-rc8 we see large amounts of oplog being read into cache and large amounts dirty data in cache and resulting operation stalls. This does not occur at the corresponding point B when running on 3.4.10.
FTDC data for the above two runs attached.