Key points are not available for this paper at this time.
The growing popularity of microservices architectures generated the need for tools that orchestrate their deployment in containerized infrastructures, such as Kubernetes. Microservices running in separate containers are packed in pods and placed in virtual machines (nodes). For applications with multiple communicating microservices, the decision of which services should be placed in the same node has a certain impact on both the running time and the operation cost of an application. The default Kubernetes scheduler is not optimal in that case. In this work, the service placement problem is treated as graph clustering. An application is modeled using a graph with nodes and edges representing communicating microservices. Graph clustering partitions the graph into clusters of microservices with high-affinity rates. Then, the microservices of each cluster are placed in the same Kubernetes node. A class of methods resorts to hard clustering (i.e., each microservice is placed in exactly one node). We advocate that graph clustering should be fuzzy to allow high-utilized microservices to run in more than one instance (i.e., pods) in different nodes. ModSoft-HP Scheduler is a custom Kubernetes scheduler that takes scheduling decisions based on the results of the ModSoft fuzzy clustering method followed by heuristic packing (HP). For proof of concept, the workloads of two applications (i.e., an e-commerce application, eShop, and an IoT architecture) are given as input to the default Kubernetes Scheduler, the Bisecting K-means, and the Heuristic First Fit (hard) clustering schedulers and to the ModSoft-HP fuzzy clustering method. The experimental results demonstrate that ModSoft-HP can achieve up to 90% reduction of egress traffic, up to 20% savings in response time, and up to 25% less hosting costs compared to service placement with the default Kubernetes Scheduler in the Google Kubernetes Engine.
Petrakis et al. (Fri,) studied this question.