Fair Online Power Capping for Emergency Handling in Multi-Tenant Cloud Data Centers

Abstract

In view of the high capital expense for scaling up power capacity to meet the escalating demand, maximizing the utilization of built capacity has become a top priority for multi-tenant data center operators, where many cloud providers house their physical servers. The traditional power provisioning guarantees a high availability, but is very costly and results in a significant capacity under-utilization. On the other hand, power oversubscription (i.e., deploying more servers than what the capacity allows) improves utilization but offers no availability guarantees due to the necessity of power reduction to handle the resulting power emergencies. Given these limitations, we propose a novel hybrid power provisioning approach, called $\mathsf {HyPP}$ HyPP , which provides a combination of two different power availabilities to tenants: capacity with a very high availability (100 percent or nearly 100 percent), plus additional capacity with a medium availability that may be unavailable for up to a certain amount during each billing period. For $\mathsf {HyPP}$ HyPP , we design an online algorithm for the operator to coordinate tenants’ power reduction at runtime when the tenants’ aggregate power demand exceeds the power capacities. Our algorithm aims at achieving long-term fairness in tenants’ power reduction (defined as the ratio of total actual power reduction by a tenant to its contracted reduction budget over a billing period). We analyze the theoretical performance of our online algorithm and derive a good competitive ratio in terms of fairness compared to the offline optimum. We also validate our algorithm through simulations under realistic settings.