AMIR

Abstract

Service demand burstiness, or serial correlations in resource service demands, has previously been shown to have an adverse impact on system performance metrics such as response time. This article proposes AMIR , an analytic framework to characterize burstiness and identify strategies to reduce its impact on performance. AMIR considers an overtake-free system model consisting of multiple queues that service multiple classes of sessions, i.e., sequences of requests. Given the per-class service demand distributions and number of sessions belonging to each class, AMIR can identify an ordering of sessions, i.e., a schedule, that minimizes burstiness at the bottleneck. Hence, it is likely to improve system responsiveness metrics, including mean session wait time and total schedule processing time. To characterize burstiness, the technique uses an order O schedule burstiness metric β O representing the mean probability that O + 1 consecutive sessions in the schedule have resource demands at the bottleneck greater than the mean bottleneck demand of the schedule. For a given O , AMIR uses Integer Linear Programming (ILP) to identify schedules that progressively minimize β i ∀ i ∈ {1, … O }. We conduct an extensive simulation study to provide insights on the conditions under which such schedules can improve system responsiveness. These results show that schedules derived from AMIR can significantly outperform those derived from baseline policies such as Shortest Job First (SJF) and random scheduling when session classes are dissimilar from one another in terms of their service demand distributions. Furthermore, minimizing for higher orders of schedule burstiness is most critical when the bottleneck is heavily utilized and when the service demands of a workload are highly variable. For the system model that we consider, we are not aware of other techniques that are designed to analytically derive insights on the performance impact of high-order service demand burstiness.