Design joint, truthful, and fair online mechanisms in colocation data centers and edge nodes

Abstract

In recent years, colocation data centers and edge nodes become important roles for providing cloud services. They usually are built in urban areas, where are closer to users, which can provide faster and more reliable services. In colocation data centers or edge nodes, users can rent a certain number of infrastructures which they just need (e.g., 100 servers for a specific service in a particular area, or 200 servers when the requirement increase), therefore, no need to build an entire data center for that particular service, which is suitable for small companies.

Although this type of infrastructure organization is flexible and economical, there are some problems to consider when selling and organizing resources in colocation DCs and edges. When the power supply to the data center fluctuates due to emergency events (extreme weathers), for ensuring the QoS, the data center operator needs to fulfill the power gap, such as using diesel generators. However, it is expensive and environmentally unfriendly. Moreover, reducing assigned power from tenants is also reasonable but still facing problems: first, the future emergency demand responses are unknown, but we need to schedule them in advance. Second, if the operators consider reducing power from batch workloads (i.e., delay-insensitive jobs) with long-term budgets, we need to make online decisions to minimize the social cost rather than greedy decisions. Third, it is meaningful to achieve a win-win solution which minimizes the operator's cost and maximizes the users' utility simultaneously. Fourth, a practical mechanism should consider not only the overall social cost but also the fairness for each user. Finally, in edge nodes, as idle resources are appearing in an online manner uncertainly, precisely predicting the idle resource and efficiently allocating them are the key to improve the utilization of edge resources.

Regarding these five problems, we design and analyze five online mechanisms, i.e., FairDR, BatchDR, joint online mechanism, HyPP, and UNEARTH, to tackle these problems systematically.

a) FairDR ensures tenants' truthfulness in the auction process, attains a bounded overall cost saving compared to the offline optimum, and guarantees strict fairness. b) BatchEDR decides the tenants' workload deferment/reduction and diesel usage upon receiving an EDR signal, for cost minimization throughout the entire EDR event, considering limited reduction budgets across multiple data centers. c) In the joint online mechanism, we propose a truthful incentive auction that provides tenants with monetary remuneration for EDR energy reduction. We seamlessly combine proposed online job scheduling algorithms for each tenant and the truthful auction, which still achieve the theoretical guarantees for both. d) In HyPP, we design an online algorithm to coordinate tenants' power reduction at runtime when power demand exceeds capacities under hybrid contracts. Our algorithm aims at achieving long-term fairness in tenants' power reduction. e) UNEARTH utilizes different types of unused resources offered by an edge computing provider. Notably, the exact amount of unused VM resources is unknown before selling them. All of our proposed mechanisms outperform the corresponding existing works and achieve performance guarantees in theoretical.