RL-Based Approach to Enhance Reliability and Efficiency in Autoscaling for Heterogeneous Edge Serverless Computing Environments

초록

Edge serverless computing represents a rapidly advancing technological paradigm with various applications across multiple computing domains. However, resource constraints and workload variability significantly impact the availability, reliability, and scalability of edge serverless systems. To mitigate these challenges, we propose the implementation of reinforcement learning (RL) to optimize dynamic autoscaling configurations within Knative for edge servers. Our research is centered on developing specialized RL environments and agents specifically designed for edge computing scenarios, considering factors such as resource limitations, network variability, and proximity to end devices. We present a system architecture incorporating an RL agent into the existing infrastructure and demonstrate its efficacy in real-world edge computing environments. The experimental results indicate that our RL-based approach outperforms manual configurations, achieving a reduction in average latency of approximately 25% (from 7-12 milliseconds to 6 milliseconds) and an increase in throughput of over 40 % (from around 150 requests to 250 requests per 300-second episode). Furthermore, our solution enhances resource utilization in terms of CPU and memory management. These findings underscore the potential of intelligent autoscaling to improve the performance, reliability, and efficiency of edge serverless applications, thereby addressing the limitations associated with static configurations. This study highlights the significant impact of dynamic reinforcement learning on enhancing the dependability, availability, and scalability of edge computing systems.

키워드