the target mongod does parallelsort from all servers, getting the records sorted.
It should just reduce and insert to temp collection.
Instead the result is emitted, which puts it into a map.
Then when all records are done, map is dumped to disk.
May be very costly in cpu and mem for large jobs