Comment by DoubleFree

The postgres query optimizer will try to minimize the number of pages read from disk (and the number of intermediate pages written to disk). Benchmarking the query optimizer by making the shared buffers large enough to hold all the data therefore seems wrong, as you're then measuring the speed of the query optimizer and the join processor, instead of the quality of the generated query plans. It would not surprise me if the generated plans for these versions are actually all the same and this is only measuring execution speed.