A two-time-scale load balancing framework for minimizing electricity bills of Internet Data Centers. Issue 5 (October 2016)

• Record Type: Journal Article
• Authors: Dou, Hui; Qi, Yong; Wei, Wei; Song, Houbing
• Is Part Of: Personal and ubiquitous computing. Volume 20:Issue 5(2016)
• Page Start: 681
• Page End: 693
• Publication Date: 2016-10
• Subjects: Cyber-physical systems -- Internet Data Center -- Electricity bills -- Two-time-scale load balancing -- On-site wind power; Mobile computing -- Periodicals; Portable computers -- Periodicals; Human-computer interaction -- Periodicals; 004.16
• DOI: 10.1007/s00779-016-0941-9
• Languages: English
• ISSNs: 1617-4909

Abstract

Abstract Internet Data Center (IDC) is one of important emerging cyber-physical systems. To guarantee the quality of service for their worldwide users, large Internet service providers usually operate multiple geographically distributed IDCs. The enormous power consumption of these data centers may lead to both huge electricity bills and considerable carbon emissions. To mitigate these problems, on-site renewable energy plants are emerging in recent years. Since the renewable energy is intermittent, greening geographical load balancing (GGLB for short) has been proposed to reduce both the electricity bills and carbon emissions by following the renewables. However, GGLB is not able to well deal with the wildly fluctuating wind power when applied into IDCs with on-site wind energy plants. It may either fail to minimize the total electricity bills or incur the costly frequent on–off switching of servers. In order to minimize the total electricity bills of geographically distributed IDCs with on-site wind energy plants, we formulate the total electricity bills minimization problem and propose a novel two-time-scale load balancing framework TLB to solve it. First, TLB models the runtime cooling efficiency for each IDC. Then it predicts the future fine-grained (e.g., 10-min) on-site wind power output at the beginning of each scheduling period (e.g., an hour). After that, TLB transforms the primal optimization problem into a typical mixed-integer linear programming problem and solves it to finally obtain the optimal scheduling policy including the open server number as well as the request routing policy. It is worth noting that the open server number of each IDC will remain the same during each scheduling period. As an application instance of TLB, we also design a two-time-scale load balancing algorithm TLB-ARMA for our experimental scenario. Evaluation results based on real-life traces show that TLB-ARMA is able to reduce the total electricity bills by as much as 12.58 % compared with the hourly executed GGLB without incurring the costly repeated on–off switching of servers.