DRLEC: Multi-agent DRL based Elasticity Control for VNF Migration in SDN/NFV Networks

Considering the fluctuations of network traffic and dynamics of unknown underlying network state, designing a elastic control model with long-term high Quality of Service (QoS) and low network cost has become a pivotal problem in Software Defined Network/Network Functions Virtualization (SDN/ NFV) network. Based on the problem, we design the multiagent Deep Reinforcement Learning based Elasticity Control approach (DRLEC). Considering multi-objective of maximizing revenue benefit and minimizing migration cost, the optimization problem for elasticity control is modeled as a Markov Decision Process (MDP). Then, taking the binary integer programming model as constraints, DRLEC is designed to solve the optimization problem of maximizing long-term profit. Experimental results demonstrate that DRLEC shows better performance than heuristics and single-agent DQN algorithm. Moreover, DRLEC can nearly achieve the upper bound of the theoretical solution, which is obtained by assuming knowing the dynamics of network traffic in advance.