This paper deals with the worst-case performance of local search algorithms for makespan minimization on parallel machines. We analyze the quality of the local optima obtained by iterative improvements over the jump, the swap, and the newly defined push neighborhood.