I would tend to second the opinion of simply choosing whatever class structure makes the most sense for the objects you need to model. While there may be some performance gain to limiting the size of the class file being read and parsed, that amount is going to be so small (unless we're talking really, really large class) that such a time saving will be lost in the overhead of disk I/O, database connections and transactions, network/internet lag, etc.
It really comes down to whether saving a few mircroseconds is worth compromising your application's design and potentially making it more difficult to maintain or reuse.
