基於邊緣多區域系統的可擴展視訊串流框架

Abstract

隨著 5G 網路的普及，對於高品質、低延遲視訊服務的需求持續上升，促使邊緣運算成為關鍵架構，藉由將資源部署於接近用戶的位置，有效降低延遲並分散傳輸負載。加上為因應使用者需求與網路條件的高度異質性，自適應串流技術被廣泛採用，其中可擴展視訊編碼（Scalable Video Coding, SVC）因具備分層調整能力而成為主流方案。然而，SVC 嚴格的層級依賴結構，加上多區域需求差異，對即時傳輸與快取資源管理造成挑戰。現有研究多聚焦於最大化使用者體驗品質（Quality of Experience, QoE），但在缺乏系統成本考量下，往往造成過度資源消耗，難以滿足服務提供者的營運需求。本研究旨在設計一套兩階段系統端優化框架，致力於在維持多區域邊緣串流服務的可接受 QoE前提下，降低整體系統成本。第一階段結合了以 ARIMA 為基礎的需求預測與粒子群最佳化 (Particle Swarm Optimization, PSO)，以主動分配頻寬與快取資源。第二階段根據目前的資源可用性使用輕量級的貪婪動態規劃 (Greedy-Dynamic Programming) 演算法調整即時傳送決策。使用真實工作負載追蹤的模擬結果顯示，我們的方法在關鍵系統指標（包括系統成本、回源率、區域負載平衡和層級滿足率）上優於基線方法。

With the widespread deployment of 5G networks, the demand for high-quality, low-latency video services has continued to rise, making edge computing a critical architectural paradigm. By placing computation and storage resources closer to end users, edge systems effectively reduce latency and distribute transmission loads. To further address the high heterogeneity in user demands and network conditions, adaptive streaming technologies have been widely adopted. Among them, Scalable Video Coding (SVC) has become a mainstream solution due to its layered scalability. However, the strict decoding dependencies inherent to SVC, coupled with regional variations in demand, pose significant challenges for real-time delivery and cache resource management. While existing studies have primarily focused on maximizing user-perceived Quality of Experience (QoE), they often neglect system-level cost considerations, leading to excessive resource consumption and limited operational feasibility for service providers. This study proposes a two-stage optimization framework that aims to reduce overall system cost while maintaining acceptable QoE for multi-region edge streaming services. In the first stage, the framework integrates ARIMA-based demand forecasting with Particle Swarm Optimization (PSO) to proactively allocate bandwidth and caching resources. In the second stage, a lightweight Greedy-Dynamic Programming (Greedy-DP) algorithm adjusts real-time delivery decisions based on current resource availability. Simulation results based on real-world workload traces show that the proposed method outperforms baseline approaches across key system metrics, including system cost, origin fallback ratio, regional load balancing, and SVC layer fulfillment rate.