Learning Framework for Virtual Network Function Instance Migration

To implement service function chains (SFCs), virtual network function instances (VNFIs) are placed on virtual machines decoupled from underlying hardwares. For cloud service providers, satisfying the quality of service of clients and reducing the resource consumption to accommodate more clients are both important. However, in the network with dynamic load, fixed VNFI placement strategies can not reach these goals. VNFI migration, which provides cloud service providers with a flexible fine-tuning strategy, becomes a potential way to solve this problem. In this paper, we formulate the VNFI migration problem with the goal to minimize the total cost in an operational cycle. Then a novel Migration-Monitor-Learning Framework (MMLF) is proposed to solve the VNFI migration problem. The most distinctive work of MMLF is that it solves the action space problem when applying reinforcement learning for VNFI migration. Action space for selection of VNFI for migration is very large and thus results in high complexity and poor convergence performance. To overcome this problem, MMLF involves a dedicated neural network to produce continuous actions under time-varying network load. Moreover, two mapping methods are proposed to map continuous actions to discrete VNFIs selected for migration. We evaluate MMLF and other state-of- the-art approaches with real trace data. The experimental results show that MMLF outperforms other approaches under various scenarios.