An analysis of Klimov's problem with parallel servers (Q1423704)

The author introduces a family of undiscounted bandits on parallel servers under controls. This family is helpful to model a variety of multi-class queueing systems including semi-Markov generalizations, machine breakdowns, priority complex patterns, etc. An index policy (called Klimov's rule) is developed. This policy is optimal for the single-server case and under some conditions, it is shown to be asymptotically optimal in a heavy traffic regime.