A discussion on readahead

It really feels like what is actually needed is to separate readahead into two separate concepts. One that exposes performance characteristics based on fixed parameters such as FS type, device type, and link type, and then another optimizer that can look at those parameters, get a feel for what kind of insane things the user is doing, and start making smarter choices based on that. That way the system can behave differently when the user is doing one big read off of a well designed local filesystem and small reads off a remote and slow filesystem.