Enhancing Interference-Aware Scheduling for Containerized Deep Learning Tasks

To efficiently manage and schedule deep learning (DL) training tasks, container technology is widely used, providing advantages such as improved scalability, resource utilization, deployment simplicity, and isolation. It is difficult to estimate resource needs for containerized DL operations with accuracy during deployment. This might possibly reduce training efficiency, especially when there is interference between shared node DL containers. This paper introduces a framework for optimizing interference in containerized DL training tasks based on DL model features, such as floating-point operations per second (FLOPs) and memory parameters. Within our framework, we assessed the resource requirements of individual DL task containers by considering model features and historical task data. We then measured interference levels between co-located DL tasks. To determine the best location, we examined the mapping quality between work nodes and containers based on their load requirements and interference level. Our experiments validate the effectiveness of our method in optimizing load balancing and reducing interference in DL clusters.