Layered Server-Client Model for Asynchronous Parallel Distributed GA and Its Parameter

K. Kojima and M. Ishigame (Japan)


Genetic Algorithm, Asynchronous, Parallel Distributed GA, Layered, Server-Client Model, Load Distribution


In Parallel Distributed Genetic Algorithm (PDGA), many studies have been done by using Coarse-grained paral lelization. Among these studies, many researches are done by using Synchronous Ring model (Stepping Stone model), Random Exchange model and Sigma Exchange model. However, there are some problems in these models: the communication does not occur efficiently in Synchronous Ring model, it is difficult to implement in Random Ex change and Sigma Exchange, and it is difficult to coordi nate a parameter in Sigma Exchange model. To solve these problems, we proposed a implementation method by us ing Server-Client model. This model consists of a server program and some client programs. A client program exe cutes GA independently and a server manages the elite in formation of each clients. To implement this model is easy and the communication occurs asynchronously. However, when we apply Server-Client model to huge search space problems, this model needs a lot of clients and high traffic loads to the server. This paper introduces a implementation method of Layered Server-Client model. This model consists of a master server program, some server program and some client program managed by a server. This means that there is a master server program above some server programs and this system may distributes traffic load. We apply Server Client model and Layered Server-Client model to TSP with 30 cities and 100 cities, and compared about the fitness value, the execution time and the traffic. From the re sult, we consider the effectiveness of Layered Server-Client Model. We also consider the effectiveness of the parame ters of Layered Server-Client Model.

Important Links:

Go Back