Collaborative Management Methods For Energy Efficiency In Edge Datacenter

“Internet of Everything” makes the computing paradigm towards distributed edge com-puting,whose core infrastructure—edge datacenters have been deployed widely in densely populated areas.The vast scale and inefficient energy efficiency management not only in-crease the energy consumption of edge datacenters,but also increase the environmental pres-sure under the goals of carbon peaking and carbon neutrality.The topic of improving the en-ergy efficiency of edge datacenters has raised serious concerns from industry and academia.The features of edge workload and deployment endow edge datacenters with unique en-ergy efficiency management abilities,i.e.,the independent energy optimization by the edge and the auxiliary energy optimization with the external infrastructure.On the one hand,the dynamic and mobile edge workload enables edge datacenters to manage their energy con-sumption using workload management,etc.,and optimize energy efficiency independently.On the other hand,edge datacenters are widely distributed in urban areas,which enables them to manage energy efficiency by taking the cross-infrastructure opportunities provided by urban infrastructures,e.g.,smart grids and district heating systems(DHSs).However,it is non-trivial for edge datacenters to manage energy efficiency.First,the relationship be-tween performance and energy consumption is complex,and the performance requirement in edge computing is strict.The tradeoff between performance and energy efficiency should be carefully balanced.Second,energy efficiency management is an online problem.The man-agement decision should be made efficiently without the whole future information while avoiding long-term loss in performance and energy efficiency.Last,existing collaborative mechanisms between edge datacenters and urban infrastructures fail to motivate the edge dat-acenter operator and service provider,limiting the potential for collaborative management for energy efficiency.To deal with these challenges,the abilities of energy consumption adjustment are thor-oughly explored.Jointly with the online optimization and market mechanism,collaborative management methods for energy efficiency in edge datacenters are designed: from the in-dependent energy optimization by the edge,achieve collaborative optimization between the performance and energy efficiency,and reduce the energy consumption of edge datacenters;from the auxiliary energy optimization with the external infrastructure,utilize the demand re-sponse and waste heat harvesting to reduce the power cost and improve the energy utilization,in terms of the “demand” and “supply” of edge datacenters’ energy resource,respectively.The independent energy optimization by the edge lays the foundation for the auxiliary en-ergy optimization with the external infrastructure,and the latter provides more space for the former method.Specifically,these solutions include the following three aspects:From the independent energy optimization by the edge,targeted at the representative edge application relying on high-powered accelerators—real-time video stream analytics,a collaborative optimization mechanism of workload and energy efficiency,Dynami E,is pro-posed.An adaptive video stream analytics framework based on deep reinforcement learn-ing(DRL)is designed,profoundly exploring the complex relationship between user experi-ence and energy consumption,and optimizing the user experience while meeting the energy limitation.A user study is conducted to analyze the effect of video context on user expe-rience preference,and a multi-objective DRL approach is designed to improve the training efficiency of the adaptive policy.Jointly with a mapping approach between video context and adaptive policy,these approaches ensure a stable and great user experience as well as efficient energy consumption limitation in dynamic video contexts.Using a video stream analytics system,the high efficiency of Dynami E in reducing energy consumption and opti-mizing user experience is verified.From the auxiliary energy optimization with the external infrastructure,a compute and electricity demand response mechanism for edge datacenters,Edge DR,is proposed.Edge DR maintains the computing resource capacity and stability of the smart grid at the computing and power peaks by workload and edge datacenter status management.For han-dling the “split incentives” hurdle,a fair economic incentive mechanism is designed to moti-vate service providers,jointly with an efficient workload management approach to ensure the quality of service(Qo S).A dynamic payment and recall mechanism is proposed to balance the tradeoff between edge datacenter operators’ short-term profit and long-term revenue.For cost control,an online algorithm is designed to decide the status of edge datacenters without any future information.Theoretical analysis shows the outstanding performance of Edge DR in competitive ratio,truthfulness,etc.Through the simulation experiments driven by a real-world demand response event,the effectiveness and efficiency of Edge DR in meet-ing edge demand response requirements,ensuring Qo S,and social operational cost control are verified.From the auxiliary energy optimization with the external infrastructure,targeted at an-other form of energy—heat,an online market and win-win incentive mechanism for edge datacenter heat harvesting,Edge Heat,is proposed,which realizes heat harvesting collab-orated with DHS.Via empirical analysis,critical issues of heat harvesting in cooperation with DHS,i.e.,feasibility,market size,and profit,are discussed.To adapt to the dynamic heat harvesting scenario,a reverse auction mechanism and an online cost-control algorithm are proposed.Edge Heat is to motivate the edge datacenter heat harvesting and optimize the long-term social operational cost.The performance of Edge Heat,such as competitive ratio,truthfulness,etc.,is verified by theoretical analysis.The simulation experiments driven by real-world DHS trace demonstrate the efficiency of Edge Heat in satisfying heat harvesting requirements,minimizing social operational cost,etc.