Dynamic CPU scheduling for QoS provisioning

Distributed, large-scale, enterprise applications are commonly supported in multi-tier data-center environments. In this paper, we study a scheduling problem for sharing CPU time in a cluster of servers among a number of enterprise customers. Such sharing is typically mandated by service differentiation requirements and QoS guarantees. Our main contribution is the formal definition of a CPU allocation/scheduling problem with respect to QoS quarantees and evaluation of scheduling policies that address the following design criteria: they have provable performance, they do not require a priori knowledge of service statistics and their overhead is adjustable. We provide the necessary mathematical framework for policies that satisfy the above criteria and evaluate proposed algorithms via theoretical analysis and extensive simulations.