Online traffic-aware linked VM placement in cloud data centers

作     者：Liwei LIN David S.L.WEI Ruhui MA Jian LI Haibing GUAN Liwei LIN;David S.L.WEI;Ruhui MA;Jian LI;Haibing GUAN

作者机构：Shanghai Key Laboratory of Scalable Computing and Systems Shanghai Jiao Tong University Department of Computer & Information Sciences Fordham University 

出 版 物：《Science China(Information Sciences)》 (中国科学:信息科学(英文版))

年 卷 期：2020年第63卷第7期

页      面：182-204页

核心收录：

学科分类：08[工学] 081201[工学-计算机系统结构] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：supported in part by National Key Research Development Program of China (Grant No. 2016YFB1000502) National Natural Science Foundation of China (Grant Nos. 61525204, 61732010) SJTU Overseas Visiting Scholars Program 

主　　题：linked VMs traffic-aware online VM placement cloud data center energy efficient 

摘      要：In cloud computing, virtual machine(VM) placement plays a crucial role in data center(DC)management, as different ways of VM placement may require different system resources. As Cisco research reveals that virtualization of DC increases traffic within the DC and causes network bandwidth to become scarce resource, recent researches have been focusing on traffic-aware VM placement. However, previous traffic-aware VM placement schemes treat the VM placement as a static process in that they do not take into account the impact of the current placement decision on the subsequent placement. In this paper,we thus propose a novel online traffic-aware VM placement scheme. Our scheme views VM placement as a context-sensitive dynamic process in that the decision of every step of the placement is made aiming at helping the subsequent steps of placement to reduce the required network bandwidth in the long run. In our scheme, we consider not only inter-VM traffic but also the bandwidth constraint of a physical machine(PM) when making a VM placement decision. To realize our objective, we put those VMs with close end time in the same or close proximity PMs so that when the VMs are terminated, one can make enough room for the future arrivals so as to not only minimize the number of active PMs but also reduce networking costs. We conduct extensive simulations to verify the superiority of our scheme in terms of networking costs and energy consumption. Simulation results show that our scheme outperforms improved-best-fit-decreasing(IBFD) scheme, a revised best-fit version that takes inter-VM traffic into account, by 30%–40% on network cost under various scenarios. Our scheme also promises 10%–25% power savings compared with IBFD.