Optimality of the last-in-first-out (LIFO) service discipline in queuing systems with real-time constraints

Consideration is given to a G/G/1 queuing system where the objective is to maximize the probability that a customer´s system time does not exceed a given deadline. The deadline is defined to be the maximum amount of time the customer can spend in the system. The authors show that for deadlines that are i.i.d. random variables with concave cumulative distribution functions, last-in-first-out (LIFO) gives the highest probability of success, and first-in-first-out (FIFO) the lowest, over the class of all work-conserving nonpreemptive service disciplines that are independent of service time and deadline. The results are extended to include policies that allow preemption in G/M/1 systems, deadlines imposed on queuing (rather than system) time, and multiple server queues