Checksums on modified pages should only be computed immediately prior to transferring pages to disk. DIRECT_IO makes that transfer point explicit so you can compute the checksum exactly once per transfer regardless of the number of page modifications. If you use the kernel cache like PostgreSQL, you must apply the checksum pessimistically because you don't know when the page will _actually_ be written back to disk by the kernel in-between page modifications.

Computing page checksums burns quite a bit of memory bandwidth, and memory bandwidth is a major resource bottleneck in modern databases. In the DIRECT_IO case, the checksum memory bandwidth should be the same as your storage bandwidth; pessimistic checksumming without DIRECT_IO can burn memory bandwidth that significantly exceeds the storage bandwidth.

Particularly in cases where you have modern, high-performance storage devices, like PCIe connected flash arrays, you can't afford to burn memory bandwidth on checksumming beyond the minimum technically required.