I think everyone agrees that you should start with doing it for a
particular FS, and then after you have done it for one, you will know
enough about what needs to be done that you can make your case that it
should be done in VFS. Frankly, I think that you should either share
caches between nodes (NUMA), or (somehow, and there are so many ways...)
divide the workload between the machines such that they don't access the
same data except in response to failure.
Hans