Several kind of services discovery mechanisms already exist to look server up in distributed systems. Each of them has its own properties and addresses specific service lookup issues. To provid access to computational servers, gathering a list of the available servers which can solve a given problem is not enough. Indeed, the characteristics of the servers have to be taken into account to optimize the global resource use. Thus, on a problem submission, a discovery service must collect from each server a prediction of the time needed to solve this problem according to its current load. An original server discovery mechanism has been developed to meet this specificities. This paper presents scalability issues when extending this service to a wide network and proposes some solutions.