Open computational systems comprise physical entities coordinating their activities in dynamic environments. Many exciting applications require a large number of such entities to achieve team coordination in complex missions execution. To meet the fundamental challenge of role allocation in such extreme teams, we propose an algorithm called LA-DCOP, that overcomes the limitations of previous algorithms by incorporating three key ideas. First, we represent the role allocation problem as a distributed constraint optimization problem and use tokens representing roles to minimize constraint violations. Second, we use probabilistic information about the team to guide the search quickly towards good solutions. Third, we designed the algorithm to manage constrained roles. We show that LA-DCOP not only meets our requirements in extreme teams, but also compares favorably against previous role allocation algorithms. LA-DCOP has allowed an order of magnitude scale-up in extreme teams, with role allocation in a fully distributed proxy-based teams with up to 200 members.