MongoDB's storage engine is more or less a mmap'd linked list of documents. It has a lot of issues once you actually start doing reads or writes, whether it's because your working set exceeds RAM or you actually want durability. It's a nice term-paper DB implementation but there's a good reason why most of the big single-instance RDBMS's use their own I/O algos instead of delegating to the kernel. Fundamentally the kernel doesn't know your optimal access pattern or your desired tradeoffs.

Redis turned away mmap'd storage some time ago; you can snapshot the DB to disk, but that's done via a fork() and the writes happen in that other process.