Performance analysis of an energy-saving strategy in cloud data centers based on a MMAP[K]/M[K]/[formula omitted] non-preemptive priority queue

Abstract

With the development of cloud computing, big data, artificial intelligence and other next generation information technologies, the scale of the data center industry worldwide is growing rapidly, resulting in a sharp increase in energy consumption. In order to efficiently reduce the idle energy consumption in cloud data centers, in this paper, we propose an energy-saving strategy based on two-threshold hysteresis cluster scheduling mechanism, which enables the reserved cluster to be dynamically switched on and off with load variation. In addition, the tasks to be processed are classified into real-time tasks and non-real-time tasks. To keep track of the two classes of tasks and model the correlated traffic in cloud data centers, we describe the arrival flow of tasks as a Marked Markovian Arrival Process (MMAP). Accordingly, we develop a non-preemptive priority queue as the system model to capture the working principle of the proposed strategy. By using the matrix-geometric solution and Gauss–Seidel method, the steady-state distribution of the system model is analyzed, and some key Quality of Service (QoS) metrics and Total Cost of Ownership (TCO) metrics are calculated. Results of numerical experiments show that, under the proposed energy-saving strategy, the overall power consumption in a small cloud data center can be reduced by an average of 30.20% under different scenarios, and the larger the cloud data center scale, the more obvious the energy saving effect. The results also confirm the impact of inter-class correlations of the input process on performance metrics. Furthermore, we identify the Pareto optimal solutions for trading off the overall power consumption and average waiting time of real-time tasks.