In this paper we verify and study through implementation our earlier game theory algorithm for load balanced dynamic channel assignment. In our algorithm every cognitive radio is initially assigned a channel in a random fashion. The radios periodically evaluate a cost function related to a Quality of Service (QoS) parameter and may change the transmission channel with probability proportional to the excess of the current cost over a cost threshold set by the user application. The actions of the cognitive radios are concurrent and the decision to switch the channel is only based on the current local link quality, which means that neither network state information exchange nor history keeping is required. The earlier theoretical work has shown that the balls-and-bins game theory algorithm has attractive convergence properties and low complexity. In this paper we evaluate the convergence properties of our dynamic channel assignment protocol over real channel conditions for different number of channels and traffic distributions. Our testbed results show that the optimal channel allocations are found in most of the cases within five to eight iterations depending on the complexity of the test case. The proposed algorithm is a good candidate for resource allocation using only local information.