Learning-based dynamic scalable load-balanced firewall as a service in network function-virtualized cloud computing environments

Abstract

Network function virtualization (NFV) is a network architecture which tries to provide communication services in clouds through virtualization techniques. Actually, NFV combines server and service and replaces a lot of network devices. NFV deploys software applications instead of hardware devices and therefore reduces network provider’s financial costs and facilities manageability. One of the services that NFVs present is virtualized firewalls in clouds. As other services in clouds, firewalls should be dynamically scaled to the needs of any business and adapt as demands increase. In this paper, a method is proposed for dynamic auto-scalability of the firewall service in cloud environments. The proposed method also balances incoming load among different virtualized firewalls which are installed as a software on virtual machines and are located in one pool. We consider a queuing model for each virtual machine. The goal here is to determine the number of active virtualized firewalls required in different time steps according to the intensity of incoming load and the proportion of total requests that goes to each firewall. Decisions are made regarding the utilization of firewall virtual machines so that QoS requirements can be met; at the same time, the resources will be saved in order to balance the performance with the cost of allocated firewall virtual machines. To solve the problem, we propose a hybrid genetic algorithm and reinforcement learning-based approach, namely GARLAS (genetic algorithm and reinforcement learning-based autonomic scaling), implemented in a cloud manager. The results of simulation with MATLAB on different realistic workloads demonstrate that the approach is able to find an optimal policy in both scalability and load balancing aspects. Also, it leads to 87.91 and 85.15% of lower average response time and 9.93 and 11.77% of improvement in utilization in comparison with static and threshold-based approaches, respectively.