I'd caution against using the filesystem unless you think things out very carefully. Typically you'll have a binary tree for the index in an RDBMS which gives you log(N) scaling on lookups for joins etc. If you use the filesystem your lookup time will probably be linear. You also pay a penalty in space for the fact that files have to be a whole number of inodes, so having many small files is a really inefficient use of disk space.

In essence, RDBMS are highly optimised for the sort of things people do with databases, but you need to know what you're doing with them.