Energy-Aware Virtual Machine Management in Inter-Datacenter Networks Over Elastic Optical Infrastructure

Liang Zhang, Tao Han, Nirwan Ansari

TL;DR

This paper tackles the challenge of minimizing energy consumption in inter-datacenter networks by intelligently migrating virtual machines (VMs) to leverage renewable energy sources. They introduce a novel "renewable energy-aware inter-datacenter VM migration" (RE-AIM) problem, formulated as a complex many-manycast communication problem, and propose several heuristic algorithms to efficiently solve it.

What Problem Does This Paper Address?

The ever-increasing demand for data processing in datacenters (DCs) leads to massive energy consumption, a significant portion of which comes from non-renewable "brown" energy sources. While renewable energy (solar, wind) offers a greener alternative, its availability fluctuates and often doesn't align with DC energy demands. This mismatch results in inefficient use of renewable energy and continued reliance on brown energy. Existing solutions for VM migration between DCs don't adequately address the dynamic nature of renewable energy availability and the complexities of modern elastic optical networks (EONs). This paper aims to bridge this gap by proposing a framework for energy-aware VM migration that optimizes the use of renewable energy and minimizes brown energy consumption within EONs.

Methods

The authors formulate the "renewable energy-aware inter-datacenter VM migration" (RE-AIM) problem. This problem is characterized by:

  • Many-to-Many Communication: Unlike traditional one-to-one or one-to-many communication, RE-AIM involves multiple source DCs potentially migrating VMs to multiple destination DCs simultaneously.
  • Integer Linear Programming (ILP): For smaller network configurations, the RE-AIM problem is formulated as an ILP, which can be solved optimally using tools like CVX and Gurobi.
  • Heuristic Algorithms: To handle the NP-hard nature of the problem for larger networks, several heuristic algorithms are proposed:
  • Anycast-SP (Shortest Path): Prioritizes finding the shortest path for VM migration.
  • Anycast-MP (Maximum Bandwidth Path): Selects paths with the most available bandwidth.
  • Anycast-EP (Ergodic Path): Uses a weighted path selection based on resource availability.
  • Anycast-JRE (Joint Resources Ergodic): A more complex algorithm that considers multiple resources for path selection.

These heuristics aim to balance VM migration requests with network capacity and renewable energy availability.

Datasets

Not discussed in the paper. The authors use simulated network topologies (six-node and NSFNET) with randomly generated user requests.

Key Insights and Ideas

  • Renewable Energy Mismatch is Key: The core problem is the temporal and geographical mismatch between renewable energy generation and DC demand, which VM migration can help alleviate.
  • EONs Add Complexity: Elastic Optical Networks, while offering high capacity, introduce new constraints (spectrum continuity, path continuity) that make VM migration routing and spectrum assignment challenging.
  • Many-to-Many is the New Normal: The paper highlights that future inter-DC communication for tasks like VM migration will increasingly be many-to-many, requiring new algorithmic approaches.
  • Heuristics are Necessary for Scale: The NP-hard nature of the RE-AIM problem necessitates efficient heuristic algorithms for practical, large-scale deployments.
  • Migration Granularity Matters: The number of migrations and the bandwidth used per migration significantly impact energy savings. More migrations, up to a point, can lead to better brown energy reduction.

Key Results

  • Significant Energy Savings: The proposed heuristic algorithms can reduce the cost of brown energy consumption by up to 19.7% compared to a strategy without migration.
  • Heuristic Performance: Anycast-EP and Anycast-JRE generally achieve performance closest to the optimal ILP solution, with Anycast-JRE often showing the lowest brown energy cost due to its comprehensive resource consideration.
  • Trade-offs: Anycast-MP, while prioritizing bandwidth, sometimes leads to fewer migrations and may not always select the best destination DC in terms of computing resources. Anycast-SP is simpler but can be less effective under heavy network load.
  • Umax Impact: Increasing the available network resource ratio (Umax) generally leads to lower brown energy costs, as more VM migrations can be accommodated.

Limitations and Future Work

  • NP-Hardness: The RE-AIM problem is proven to be NP-hard, meaning finding the absolute optimal solution for large networks is computationally infeasible.
  • Heuristic vs. Optimal: While heuristics perform well, there's still a gap between their results and the optimal ILP solution for smaller instances.
  • Future Work: The authors suggest extending the work to consider multiple VMs belonging to the same application being migrated together, and to incorporate more sophisticated server selection within DCs. They also mention the potential for further optimization by considering dynamic changes in renewable energy availability and traffic load.

Summary updated

BibTeX
@article{zhang2018energy,
  author = {Liang Zhang and Tao Han and Nirwan Ansari},
  title = {{Energy-Aware Virtual Machine Management in Inter-Datacenter Networks Over Elastic Optical Infrastructure}},
  journal = {IEEE Trans. Green Commun. Netw.},
  year = {2018},
  doi = {10.1109/tgcn.2017.2771724},
}
Figure