Energy-Aware Inter-Data Center Virtual Machine Migration over Elastic Optical Networks




Fatima, Salehnejad Amri

Journal Title

Journal ISSN

Volume Title



The rapid growth of data processing demands in large-scale data centers (DCs) has increased the network's brown energy (BE) consumption. The BE is generated from fossil fuels and has an adverse effect on the environment. Since most DCs are now powered by both BE and renewable energy (RE), migrating workloads from DCs with insufficient RE to DCs with sufficient RE can decrease the total BE consumption in the network. However, selecting a destination DC under the dynamic nature of the underlying network and without any prior information is challenging. In addition, although migration can help reduce BE consumption, it comes with an additional cost due to using network devices for the migration. This thesis focuses on minimizing the total cost, which includes BE consumption costs, migration costs, and optical network device costs. To achieve this goal, we optimize both the DC selection process and the efficient transfer of virtual machines (VMs) between DCs. In Chapter 3, we formulate the DC selection as a two-stage Multi-Armed Bandit (MAB) problem. First, we define an arm as a destination DC, and select a DC with the lowest power consumption and available RE. In the second stage, we define the arm as a path, and we implement MAB to find a path with the lowest delay from the source to the selected destination DC. Proposing and utilizing the modified sliding-window lower confidence bound (MSW-LCB), we estimate the lowest power consumption among DCs and the lowest migration cost at each round to find a proper destination DC and path, respectively. Additionally, we adopt optical grooming techniques to minimize the cost of optical network devices used during VM transfer. Furthermore, to validate the effectiveness of our algorithm, we conduct an evaluation using three different real-world datasets to provide different inputs for the algorithm. This evaluation is assessed on USNET topology. In comparison to the sliding-window lower confidence bound (SW-LCB) and two other MAB-based algorithms, namely the knapsack–based upper confidence bound (KUBE) and $\epsilon$-Greedy, the MSW-LCB approach reduces the total cost by about 15\%, 23\%, and 34\%, respectively, while having low regret. The MSW-LCB regret and migration costs are demonstrated to be around 15\% lower than SW-LCB, respectively. We also evaluate our algorithm in two scenarios—one with and one without optical grooming—to show the efficiency of optical grooming in decreasing optical network costs. The results indicate a 12\% drop in network costs. The results of this evaluation provide valuable insights. This indicates that our algorithm is capable of handling real-world data and effectively addressing the challenges associated with inter-DC VM migration.



Multi-Armed Bandit, VM Migration, Optical Grooming, Elastic Optical Networks