# EPRS: Experience-Prioritized Reinforcement Scheduler in Edge Clusters

## TL;DR

This paper introduces EPRS, a new scheduling method for edge computing that improves resource balance and efficiency using reinforcement learning.

## Contribution

EPRS introduces a priority-driven sample selection mechanism for focused learning in edge cluster scheduling.

## Key findings

- EPRS achieves 28.25% better performance than existing reinforcement learning-based scheduling methods.
- The method improves load balancing by 29.78% compared to traditional scheduling algorithms.
- The framework considers multiple node-level metrics and task-specific requirements for resource allocation.

## Abstract

Edge computing has garnered significant attention in recent years due to its potential in distributed systems. However, the dynamic and heterogeneous nature of edge environments introduces substantial challenges for task scheduling. Conventional rule-based scheduling algorithms often fail to adapt to rapid load fluctuations, resulting in cluster load imbalance and suboptimal resource utilization. To address this issue, we propose a container-based edge cluster scheduling framework designed to enhance load balancing. Within this framework, we introduce an Experience-Prioritized Reinforcement Scheduler (EPRS), which leverages a priority-driven sample selection mechanism to facilitate focused learning of high-value samples. The EPRS dynamically monitors node resource states via a real-time resource monitor and optimizes multi-dimensional resource allocation by jointly considering node-level metrics (e.g., computational resources, memory pressure, storage performance, and container density) and task-specific resource requirements. To validate our approach, we implemented a system prototype integrated with the proposed framework and EPRS in a Kubernetes-based edge cluster. Experimental results demonstrate that the proposed method significantly improves multi-dimensional load balancing performance, achieving an average gain of 28.25% over existing reinforcement learning-based scheduling approaches and a 29.78% improvement compared with the traditional scheduling algorithm.