I have been using SQLite on network storage using Ceph replication (similar to the AWS block storage thingy). In which only one client gets access to the networked desk. And the mmap performance was exactly where you expect it to be. Reads were happening mostly from memory as the read blocks get cached (and for the record, if the SQLite connection cache is large enough to cover most of the working set then you will not see improvement with mmap, except for the case when you get frequent updates to the database from other connections, which invalidates the SQLite connection cache AFAIK) I would expect NFS (and AWS' EFS) to exhibit a different read behaviour due to the possibility of some client changing the data that is cached by some other client.