> Would you expect a file in ramdisk, perhaps just after a copy to /dev/null, WHEN OPENED WITH PRAGMA MMAP, to be near-identical speed to the malloc() (ie :memory: followed by "restore from") No... > because it goes through identical Sqlite code and accesses identical in-physical-RAM memory? Because that's not my expectation of the API. However... > Note - this is a question about Sqlite - :memory: vs mmap pointer to file ON RAMDISK. I think it is on-topic, as it is about Sqlite internals. I want to confirm the identical codepath. ie identical xFetch or whatever calls. We're going to have to wait on someone who's familiar with what sqlite *actually* does when the mmap pragma is enabled to enlighten us, but my outsider's guess is that the code paths do not become identical once mmap is enabled. Even if the paths do become identical: as Keith pointed out, mmap does not necessarily eliminate I/O. OS-level file content caching is a "black box" feature applications cannot rely on. It might work as you intend on any given day on any given machine and it might not. That is two or three layers removed from where sqlite has any influence.