On Arbitrating The Power-performance Tradeoff In SaaS Clouds

With rapid development and widespread applications of cloud computing over recentyears, not only the quantity but also the scale of datacenters worldwide have increased dra-matically. However, this also incurs skyrocketing power consumptions of such large-scaledatacenters, which consist of massive farms of servers and abundant computing availability.Under the global context of severe energy starvation and adverse environmental impact, theissue of how to balance the power-performance tradeoff of datacenters hosting diverse ap-plications has drawn substantial attention from governments, industries as well as academia.An analytical framework is presented for characterizing and optimizing the power-performance tradeoff in Software-as-a-Service (SaaS) cloud platforms. The objectives aretwo-fold:(1) Maximizing the operating revenue when serving heterogeneous SaaS appli-cations with unpredictable user requests, and (2) Minimizing the power consumption whenprocessing user requests. To achieve these objectives, a unified profit-maximizing objectiveis constructed to jointly consider revenue and cost in an economic view. An offline solutionto maximize the supreme bound of the objective is first developed, in order to (1) justify thevalidity of our theoretical model, and (2) establish a benchmark to examine the effective-ness of other control solutions. As a highlight of our contributions, we take advantage ofLyapunov Optimization techniques to design and analyze an optimal yet practical controlframework, which makes online decisions on request admission control, routing, and virtualmachine (VMs) scheduling.The proposed online control framework can be extended to accommodate a varietyof design choices and operational requirements in a datacenter. Specifically, bufferingfacilities can be introduced to alleviate the bursty admitted requests and to improve therobustness of the system, and a power budget can be enforced to improve the datacenterperformance (dollar) per watt. Mathematical analyses and simulations have demonstratedboth the optimality (in terms of a cost-effective power-performance tradeoff) and systemstability (in terms of robustness and adaptivity to time-varying and bursty user requests)achieved by the proposed control framework.