Cross-Layer Power Management For Highly-Available/-Scalable Data Centers

Compared to the good scalability of IT equipment,non-IT infrastructure(e.g.,power distribution facilities)is difficult to scale and upgrade in data centers.In this case,server oversubscription has become the primary solution to squeeze more IT equipment into data centers without upgrading existing non-IT infrastructure in order to meet users' fast-growing demand on IT resources.Nevertheless,in the presence of irregular users(e.g.,power attackers or emerging applications),this design may affect the normal operation of IT infrastructure,degrade the quality of service for regular users,which would ultimately reduce the availability and scalability of data centers.To address these issues,this paper investigates how to improve the resilience of power management in oversubscribed data centers,which requires to enhance their ability of handling abnormal power peaks and improve their flexibility in embracing emerging applications.Specifically,this paper not only studies abnormal power behaviors in traditional Iaa S/Paa S/Saa S cloud services and the possible solutions,but also actively explores the demand and challenge of power management under emerging Faa S cloud service.Totally,this paper includes the following four aspects.Heterogeneous Backup Power Architecture Based on Hybrid Pooling for Energy Storage.This architecture is designed to balance the peak shaving status of energy storage devices.It aims at preventing malicious users from causing larger server outages and even the entire data centers crash by depleting local energy storage,ultimately severely reducing system availability.The architecture deploys an integrated power management method based on both software and hardware mechanisms.First,the architecture conceals the usage information of batteries by defining a virtual battery pool,which can interfere with the attacker's identification of weak batteries.Second,the architecture deploys a small supercapacitor on each server cabinet to handle invisible power peaks.Finally,the architecture is able to proactively detect invisible attack power peaks.Therefore,it can implement precise throttling to reduce the performance loss of power capping.Incentive-Driven User Power Allocation Approach.This resource mapping approach is proposed to incentivize data centers to maximize their equilibrium of power supply and demand.It promotes the communication between power management systems and users.It can also effectively deter a malicious user from degrading the Qo S of normal users through abusing power resources.To this end,it leverages a market-based method to establish multiple communication channels between power management layer and users on exchanging the power supply and demand.In the market,different users iteratively adjust their power demand according to the expected performance-cost ratio in the bidding process.Meanwhile,the scheduling layer globally determines the allocated power for each user and the cost in the purchasing process.Besides,it employs an incentive-based reward and punishment mechanism effectively to defend against the power grab attack.Request-Aware Traffic Power Allocation Strategies.The mechanism aims to resolve conflicts and contradictions between the increasing scarcity of power resources and the gradual expansion of web application users.In an oversubscribed data center,an external web user can generate malicious peak power inside the data center by sending attack requests from the network without being detected by the network security system.The attack purpose is to affect the execution of legitimate users by triggering various power capping mechanisms.By collaborating with the network load balancer,the proposed mechanism can improve the ability of power management in handling unexpected power anomalies.On the one hand,the mechanism identifies the potentially malicious user requests at the network load balancer side and redirects them to the isolated servers.On the other hand,the power management system employs differentiated power schemes to satisfy the demand of legitimate users whilst suppress resource-hungry requests.Agile,Macro-Micro Collaborative Power Management System.The system intends to coordinate the power management modules of data centers from upper cloud-level system to bottom server-level component.The system can not only enhance data center's availability under the above power anomalies,but also agilely management power of emerging microservices.It can improve the flexibility of power management at a finer application granularity.The main feature of the system is to eliminate the conventional power management latency through converging the power monitoring system at the data center level and the power regulation module of the local server.On the one side,the accomplished system can reduce upper control latency by proactively determining power allocation for each microservice rather than iteratively comparing power usage with overall budget.On the other side,it can also eliminate bottom execution latency through enabling nanosecond-scale power scaling.Overall,this paper explores the power management architecture,power resource mapping methods and management mechanisms at different levels of the data center.It finally proposes a cross-layer power management system that combines hardware and software collaboration,architecture design and system optimization,which can ensure the availability and scalability of the data center in dealing with various cloud user power demands.