Traffic-aware Threshold Adjustment for NFV Scaling using DDPG

Abstract

Current solutions mostly focus on how to predict traffic, rather than observing traffic characteristics in a specific NFV scenario. So, most of them use a uniform threshold to scale in/out. In real NFV scenario, each VNF may serve the one or more flows, and the characteristics of these flows are completely different, a uniform threshold used in this scenario is not suitable, because each VNF has a distinct processing logic depending on incident network traffic and events. Even if certain VNFs share packet processing functionality such as packet header analysis, the differences in upper-layer processing and implementation can exhibit unique resource usage patterns. We proposes a dynamic threshold scaling mechanism that can tailor thresholds according to each VNF's characteristic. As setting thresholds is a per-VNF task, and requires a deep understanding of workload trends and the diversity of each VNF, so we have added tailor-made features to the traditional dynamic mechanism. Besides, we also reserve resources by predicting workload and add an emergency module to cope with anomaly traffic, that is to say we develop a hybrid scaling policy combining proactive and reactive scaling together. Moreover, the sharp rise of network traffic not only can be caused by large amount of new incoming flows, but also can be induced by the growing of existing flows. If the traffic arises mainly due to the growing of existing flows, then only rerouting new flows can not alleviate the overload quickly and SLAs may be violated \cite{zhang2016co}. The only method to avoid SLA violations is to migrate flows and associated NF internal states quickly and safely from existing instances to new scaled instances, so state migration is an important part of the scaling procedure. We achieved the flow migration in scaling process on openNF to guarantee the accuracy and timeline of scaling.