Towards an Optimized Containerization of HPC Job Schedulers Based on Namespaces

Abstract

Recently, container technology is gaining increasing attention and has become an alternative to the traditional virtual machines artifact. The technology is used to deploy large-scale applications in several areas such as Big Data, AI, and High-Performance Computing (HPC). In the HPC field, several management tools exist as Slurm in one hand. On the other hand, the literature has considered many container scheduling strategies. The majority of container scheduling strategies don’t think about the amount of data transmitted between containers. This paper presents a new container scheduling strategy that automatically groups containers that belong to the same group (Namespace) on the same node. In brief, the plan is application-aware as long as someone knows which containers should be grouped in the same Namespace. The objective is to compact the nodes with containers of the same group to reduce the number of nodes used, the communication inter-node costs, and improve containerized applications’ overall Quality-of-Service (QoS). Our proposed strategy is implemented under the Kubernetes framework. Experiments demonstrate the potential of our strategy under different scenarios. Most importantly, we show first that cohabitation between our new scheduling strategy and the default Kubernetes strategy is possible and for the benefit of the system. Thanks to Namespaces, cohabitation is not limited to two methods for scheduling either batch jobs or online services. Second, thanks to the deployment automation, we also demonstrate that multiple Slurm clusters can be instantiated from a pool of bare metal nodes. This reality contributes to the concept of “HPC as a Service.”