Currently DBApiLayer iterates through documents TWICE in case '_id's have to be created ( "shouldApply" is true ): once to set "_id"s and the second time to create batch chunks.
This slows down the insert ( with any WriteConcern ).

Attaching a patch that optimizes the INSERT: now documents are iterated only ONCE without changing a current behavior.

/Anatoly

P.S. I am currently testing with application level partitioning, so different partitions actually sent to the different mongoDs in real time ( even distribution ): https://github.com/anatoly-polinsky/mongodb-write-performance-playground and the patch does help to boost the performance.