We present a scheme for efficiently administrating a set of distributed simulation servers which attend requests from a number of users. These can be either sequential or parallel simulation models, or random combination of the two.We treat each server as a BSP machine and organize our scheduling strategy upon the cost prediction and computation model supported by it. Empirical results show that the proposed strategy is able to archieve near optimal performance . Our aim was to develop a mechanism which be able to work with minimun global information so that it can be utilized on Web based simulation systems.