ML-Assistant Service Function Chaining Workload Scheduler for Cloud-Native inFrastructure

Service Function Chaining (SFC) has the potential to revolutionize cloud computing and containerization paradigm. SFC jobs, predominantly deployed as containers within cloud-native infrastructures, require substantial computing, memory, and storage resources to operate Virtual Network Functions (VNFs) efficiently. However, scheduling SFC workloads in cloud-native infrastructures, such as Kubernetes clusters spanning geographically distributed data centers, presents significant chal-lenges. These challenges arise not only from the diverse scheduling strategies employed by each data center but also from the vast number of jobs processed on a daily basis. Furthermore, SFC tasks have stringent requirements for computing resources and network bandwidth, which complicates the scheduling process. This paper introduces a machine learning-based (ML-based) scheduling approach to address these challenges by employing the Long Short Term Memory (LSTM) model. This paper makes two main contributions. First, we formulate the dynamic SFC mapping problem by modeling SFC jobs in terms of computing and memory resources. Then, we use a decentralized multiagent LSTM framework to predict the hardware resource usages of nodes for each incoming SFC job, in order to select the most suitable node to optimize system resource utilization and minimize the job wait time. Our results demonstrate a 6%-20% improvement in scheduling efficiency and accuracy compared to non-machine learning-based schedulers.