On Data Center Demand Response: A Cloud Federation Approach

Monireh Mohebbi Moghaddam, Mohammad Hossein Manshaei, Walid Saad, Maziar GoudarziIEEE Access 2019 · OpenReview

Abstract

The significantly high energy consumption of data centers constitutes a major load on the smart power grid. Data center demand response (DR) is a promising solution to incentivize cloud providers to adapt their consumption to grid conditions, mitigating stability issues and reducing electricity costs. This paper proposes a cooperative cloud federation framework where independent providers collaboratively manage workloads. By leveraging cooperative game theory, specifically hedonic coalitional games, the framework enables spatial workload migration to exploit location-dependent DR incentives. Simulation results using real-world traces demonstrate that this cooperative approach increases provider profits by up to 70% and enhances smart grid performance compared to non-cooperative strategies.

Introduction

Data centers have become energy-hungry giants, placing substantial strain on local power grids. To manage this load, utilities implement Demand Response (DR) programs, offering financial rebates to consumers who reduce or shift their power usage during peak periods. While large-scale providers with multiple locations can easily participate, smaller, independent cloud providers (CPs) often lack the operational flexibility to respond effectively on their own.

Existing research typically assumes a single entity managing multiple data centers. However, the real-world landscape is comprised of multiple "selfish" and independent providers. This work addresses the challenge of how these independent entities can cooperate to maximize collective rebates and minimize energy costs without a central authority, and how such cooperation impacts the overall stability of the smart grid.

The Proposed Cloud Federation Framework

The core of this research is the "BidCoop" framework, which facilitates interaction between independent CPs and smart grid utilities. The system architecture involves four primary stages: bidding, coalition formation, workload distribution, and the communication of winning bids.

System architecture illustrating the interaction between smart grids and cloud providers through bidding, coalition formation, and workload distribution.

1. Federation Formation

The authors utilize Hedonic Coalitional Games to model the formation of cloud federations. Unlike top-down approaches, this distributed algorithm allows CPs to autonomously decide to join or leave a federation based on their own profit-seeking preferences. The algorithm is designed to reach a "Nash-stable" state, where no provider has an incentive to move.

2. Optimal Workload Allocation

Once a federation is established, members solve a Mixed-Integer Linear Programming (MILP) problem. This optimization determines the most profitable way to distribute virtual machine (VM) tasks across the combined data centers of the federation, taking into account local electricity prices, migration costs, and grid rebates.

3. Fair Payoff Distribution

To maintain cooperation among independent entities, the framework employs the Shapley Value. This game-theoretic method ensures that the total surplus generated by the federation is distributed fairly, rewarding each member based on their actual marginal contribution to the group's success.

Experimental Evaluation

The framework was evaluated using three distinct workload scenarios to ensure robustness:

  • Microsoft Azure Trace: A real-world dataset covering six days of VM arrival rates and lifetimes.
  • HP Request Trace: A high-variability one-day trace used as a "stress test."
  • Synthetic Workload: A two-day load for baseline functional testing.

Electricity pricing and migration costs (modeled at 0.001 $/GB) were based on Amazon EC2 standards. The study compared the proposed BidCoop method against a non-cooperative baseline (NoCoop) and a strategy focused solely on energy efficiency (Coop).

Key Results and Insights

The simulation results highlight several critical advantages of the federation approach:

  • Profitability: The BidCoop approach consistently outperformed other strategies, showing an average profit improvement of up to 70% over the non-cooperative case. This proves that being "grid-aware" through cooperation is significantly more lucrative than individual optimization.
  • Grid Stability: A "win-win" scenario was observed for the smart grid. By chasing rebates, federated CPs naturally reduced the Power Demand Deficit (PDD)—the gap between actual and desired power consumption—leading to improved grid stability across all involved utilities.
  • Scalability and Convergence: Despite the mathematical complexity of partitioning providers, the distributed hedonic algorithm converged rapidly, typically in fewer than 15 iterations for a 6-provider setup.
  • Resource Contribution: Analysis showed that CPs with larger data center capacities contributed more to the federation's success and were appropriately rewarded with higher profit shares via the Shapley Value.

Conclusion

This work demonstrates that cloud federation is a viable and highly effective strategy for data center demand response. By cooperating, independent providers can unlock significant financial benefits while supporting the operational stability of the smart grid. Future research will focus on integrating the high variability of on-site renewable energy sources and exploring overlapping coalitions where providers can split resources across multiple groups.

Summary updated

Abstract

The significantly high energy consumption of data centers constitutes a major load on the smart power grid. Data center demand response is a promising solution to incentivize the cloud providers to adapt their consumption to the power grid conditions. These policies not only mitigate the operational stability issues of the smart grid but also potentially decrease the electricity bills of cloud providers. Cloud providers can improve their contribution and reduce their energy cost by collaboratively managing their workload. Through cooperation in the form of cloud federations, providers can spatially migrate their workload to better utilize the benefits provided by demand response schemes over multiple locations. To this end, this work considers an interaction system between the independent cloud providers and the corresponding smart grid utilities in the context of a demand response program. Leveraging the cooperative game theory, this paper presents a federation formation among the cloud providers in the presence of a location-dependent demand response program. A distributed algorithm that is coupled with an optimal workload allocation problem is applied. The effect of the federation's formation on the clouds' profits and on the smart grid performance is analyzed through simulation. Simulation results show that cooperation increases the clouds' profits as well as the smart grid performance compared to the noncooperative case.
BibTeX
@inproceedings{moghaddam2019data,
  author = {Monireh Mohebbi Moghaddam and Mohammad Hossein Manshaei and Walid Saad and Maziar Goudarzi},
  title = {{On Data Center Demand Response: A Cloud Federation Approach}},
  booktitle = {IEEE Access},
  year = {2019},
  url = {https://openreview.net/forum?id=uR1cFoufGk},
}
Figure