Energy and Migration Cost Minimization in Virtual Network Embedding for Disaggregated Cloud Data Centers

Cloud data centers are evolving toward a disaggregated architecture to achieve higher efficiency in handling unprecedented large-scale workloads, particularly acceleratorassisted workloads such as AI training and inference. The core idea of this architecture is to decouple resources-including CPUs, memory, NICs, and accelerators-that were traditionally integrated within standalone servers, and reorganize them into shared resource pools. In this work, we study the virtual network embedding problem, which has been extensively explored in traditional server-based data centers, within the context of disaggregated data centers. Specifically, we address the consolidation of ongoing virtual network requests by migrating them to fewer nodes and links, thereby improving overall resource efficiency. We explicitly account for migration costs in the optimization process and aim to minimize them. Finally, we present Pareto fronts that characterize the trade-offs between power consumption and migration cost.