Currently the join cost model's usage of Mackert Lohman formula for estimation of IOs only considers the size of the WT page cache. In reality, MongoDB also caches pages in the OS page cache as well, making the effective size of the buffer cache larger, though potentially storing duplicates.

We could improve our estimate of INLJ/Index scans for collections which result sets which don't fit in memory by considering the effect of the addition cache.