It's _easier_ to write IO with mmap, but hand-optimized file IO could do even faster. And DBMSes historically cared a lot about optimizing file access.

So I'd expect SQLite to be faster without mmap, as I expect their developers to nurture the file access, instead of relying on OS-provided mmap.

PS: For example, say your code has a byte array, memory-mapped from a file. If the code needs to do a random read from it, it has no way of knowing whether that read would require waiting for a page to be read from disk, or it's already cached. Hand-optimized file IO have an option to maybe do some other things instead of waiting for disk (or during, or before).