Energy-Efficient and Interference-Aware VNF Placement with Deep Reinforcement Learning
Yanyan Mu, Lei Wang, Jin Zhao
Abstract
By decoupling network functions from the underlying dedicated hardware, network function virtualization (NFV) has become a promising paradigm to reduce network operating expenses. NFV can provide elastic placement of Virtual Network Functions (VNFs) in the underlying data centers. However, the co-located VNFs on the same server may suffer from performance interference due to computing-resource and memory-resource sharing. This article focuses on how to ensure the performance of each VNF while minimizing the total energy consumption of the data center. By showing that the bin-packing problem is polynomial-time reducible to our model, we prove that the offline version of this problem is NP-complete. Then, for a homogeneous environment where all servers are of the same type, we design First-Fit Heuristic (FFH) algorithm and analyze the approximation performance of it by proving the lower bound value. For the heterogeneous environments, we propose an efficient solution based on deep reinforcement learning (DRL) named DDAP (Deep Deterministic Automatic Placement). Our experiments show that DDAP can quickly respond to each request and achieve better performance. In particular, DDAP can reduce energy consumption by 7.6% and running time cost by 63.2% on average compared to state-of-the-art methods.