Parallel Genetic Algorithms (PGAs) suffer of many methodological deficiencies that preclude a unified study. Distributed and cellular GAs are the two separate approaches people consider for parallelization. However we think they two are the expression of a common underlying parallel model and that their combination could be very useful. Also a great deal of topological issues can be drawn for any problem and no guide exists on using them. Finally we think that the traditional imperative implementation for (P)GAs is old-fashioned and that the more adequate and flexible object oriented (OO) methodology should be used. We present a study on topologies, granularity and object oriented design of PGAs. We evaluate an OO system for training neural networks and another OO system for studying how PGAs of different topologies and grains perform on a set of functions and TSP’s. We conclude that cellular GAs are well suited for numeric function optimization, distributed GAs for problems with a time-consuming evaluation function and also that some combinations of different topologies and grains are more efficient than traditional PGAs.