Efficient And Fair Energy And Performance Management Of Data Center

As the backbone of cloud computing,data centers(DCs)are spending massive electricity costs annually.It is reported that the total energy cost of DCs within US will reach$13 billion in 2020.Therefore,DCs are paying more attentions to energy management,which can be divided into two categories: management of energy supply and demand response(DR).Specifically,management of energy supply means using cheaper energy sources(e.g.,renewable energy)to power up DCs,while DR means reducing DCs' energy bills via adjusting their power consumptions in time and space according to electric rates.In this dissertation,we first study how to reduce DC's energy bill via energy supply management and DR.However,focusing on reducing DC's energy cost may cause performance degradation or increase of other costs,e.g.,IP transit cost.To solve this problem,we propose a mechanism to balance DC's costs and performance.To further improve the efficiency of DC energy management,we design a novel pricing policy to incentivize users of DC cooperate with DC to reduce energy bill.In conclusion,these works can be divided into the following four topics.Some related works suppose that DCs can make precise predictions on their energy demands,which are impractical.In fact,DCs have limited predication capabilities,and incorrect predictions may cause ominous ramifications.For instance,overestimation of energy demand can lead to waste of money,which is used to build the electric system,and underestimation of energy demand may cause undersupply of energy.In this dissertation,we explore how to realize cost-effective self-powered DCs with limited prediction capabilities via energy management.Here,we utilize renewable energy,fuel cells and rechargeable batteries simultaneously to power up DCs.Considering that the DCs have limited prediction capabilities,we propose online mechanism about management of energy supply.With this mechanism,self-powered DCs can effectively reduce their energy bills based on the short-term predictions of energy demands and renewable energy supplies.Some related works aim at reducing DCs' energy consumptions,which may be useless for reducing DCs' energy bills when the DCs confront complex electric rates.For instance,a DC's energy bill is decided based on its peak power.In this case,to reduce energy bill,one should smooth the DC's power instead of reducing its overall energy consumption.In this dissertation,we explore how to reduce DCs' energy bills with DR strategies when confronts complex electric rates.Meanwhile,considering that many jobs executed in DCs are sensitive to performance,we ensure that the quality of service(QoS)is always tolerable for each job during energy management.With the mechanism proposed in this dissertation,DCs can deduce suitable DR strategies based on electric rates and reduce their energy bills effectively via shedding or time-shifting their energy consumptions.Some related works can effectively reduce DCs' energy bills at the expense of performance degradation or increase of other costs of DCs.For instance,shifting loads of one DC to another,which confronts cheaper electricity price,can increase the amount of data transferring between the two DCs and the corresponding IP transit fee.To overcome this problem,we first propose a tractable mathematical expression to model DC's IP transit cost and several strategies for management of data transferring.Next,we design a mechanism to balance a DC's costs and performance.With these works,a DC can effectively reduce its IP transit cost and achieve a suitable trade-off among energy bill,performance and IP transit cost.Most related works do energy management with limited flexibility of load adjustment,which indicates limited flexibility of energy management.For instance,a DC hosts many science computing jobs,which have stringent deadline requirements.In this case,to ensure QoS,the DC cannot postpone any job for energy management.To overcome this difficulty,we design a non-intrusive pricing to provide users of DC with monetary incentives,which can incentivize the users to loosen their requirements of QoS.In this case,we can obtain win-win between DCs and their users.With this pricing policy,DCs can obtain higher administrative authorities for load managements,and thus reduce their energy bills more effectively.For users,they can obtain the capabilities of making trade-off between costs and performances with the pricing policy.Last but not least,considering that the DCs have limited capabilities of load predictions and are not informed about the users' private informations,we propose online update of energy management mechanism.With our design,DCs can effectively and fairly manage their energy consumptions and performances even when they have made inaccurate predictions about their loads and the users' QoS requirements.