Dynamic Decomposition and Deployment of the Virtual Network Functions with Microservices using Deep Reinforcement Learning

The 6G network is anticipated for applications and services with faster data rates, ultra-reliability, and lower latency than 5G, which will enable the IoT to expand further. These highly demanding 6G applications will burden the network by imposing stringent performance requirements. NFV is a novel paradigm aiming to minimize operational and capital costs by decoupling network functions from their dedicated proprietary and allowing them to run as softwarized NFs (say, Virtual Network Functions (VNFs)) on commodity hardware. Although NFV presents a promising solution, it also poses substantial Resource Allocation (RA) challenges, classified as an NP-Hard problem. To solve this problem, we investigated the potential of DRL approach for deploying VNFs under several QoS constraints. More importantly, the failure recovery requirements are focused on the node-outage problem, where an outage can be due to a disaster or unavailability of network topology information. Our model provided almost 70–90% Service Acceptance Rate (SAR) even with a 50% nodal outage for certain combinations of scenarios.
6G network services will be immensely complicated and relatively short-lived; network operators will be forced to deploy them flexibly, on-demand, and agilely. Microservice technique is proposed to address the aforementioned concerns, in which the services are loosely decomposed, resulting in enhanced deployment flexibility and modularity. This research looks into a novel RA technique for microservices-based NFV that allows for the faster and more dynamic deployment and decomposition of VNFs. VNF decomposition introduces additional overheads that have a negative impact on network; thus, finding the correct balance of when and how much decomposition to allow is crucial. Therefore, we proposed a criterion for determining the potential/candidate VNFs for decomposition and the granularity of such decomposition and we constructed a DDQL model for deployment. This demonstrated that the microservices method had a nearly 50% higher normalized SAR than the monolithic deployment of arriving services.
Traditionally, the deployment methods, like priority-based, revealed the existence of a starvation problem. Starvation means longer waiting times and eventual rejection of low-priority services due to a ‘time out’. To overcome this, an ‘Adaptive Scheduling’ (AdSch) model is proposed, proved to be more reasonable than traditional methods. Moreover, a ‘Dynamic Prioritization’ (DyPr) allocation method is also proposed for unseen services, and the importance of macro- and micro-level priority is studied. We presented a zero-touch solution using DDPG for AdSch and an online-Ridge Regression (RR) model for DyPr. The DDPG successfully identified the ‘Beneficial and Starving’ services, efficiently deploying twice as many low-priority services as others, reducing the starvation problem. Our online-RR model learns the pattern in less than 100 transitions, and the prediction model has an accuracy rate of more than 80%. The starvation problem is explored not only based on the QoS but also in the realm of traffic. We delved deep into the traffic generated by the unseen services to understand them better. The goal was to differentiate the highly demanding service or, say, high revenue-generating services from the not-so-demanding service. The trained cluster+classifier model-- Aggo+DT categorizes online SFC services as HD or NHD. Increasing the scheduling and placement model's awareness of high-revenue-generating services. As a result, a trade-off occurs between low-priority HD SFCs and low-priority NHD SFCs is seen, resulting in a nearly-suboptimal resource allocation solution. This traffic-aware model was able to deploy twice as many low-priority HD services as the non-traffic-aware model, proving that our proposed model successfully identified and favoured the Beneficial and Starving high-revenue generating services for deployment.