- 學位論文
第五代行動通訊雲端無線接取網路之動態資源管理與運營最佳化
Dynamic Resource Management and Operation Optimizations for 5G Cloud Radio Access Networks
摘要
第五代行動通訊(5G)的資源管理效率對於無線通訊技術發展有著重要的影響，起因於系統架構中採取將基頻單元(Base Band Unit, BBU)分離至後端雲端無線電接取網路(Cloud Radio Access Network, C-RAN)伺服器中，並以虛擬化(virtualization)技術將BBU所需的運算工作(task)集中於基頻池(BBU pool)進行管理，而管理核心除了負責控制該BBU的允入、配置、轉移與回收為主要工作之外，另須控制伺服器的開關機狀態以擴增或縮減運算空間與能力；因此本論文主要研究議題是基於增強式學習之系統架構設計有效率的管理核心機制演算法與運營伺服器策略達到系統資源最佳化為目的。研究範圍主要根據C-RAN的網路規劃與運營階段，以通訊(網路)及運算(系統)角度，運用不同效能指標與解題程序所形成之相關議題列述如下： 第3章：基於馬可夫決策過程之C-RAN動態資源管理；面對需求的即時性(on demand)及延展性(scalability)，如何得知長期運轉後之BBU允入與暫時性阻斷策略？本議題應用馬可夫決策過程，針對BBU個別到達速率不同，發展在系統的各個狀態下運算該需求之最佳決策以最大化服務效能為目標。 第4章：最佳化運算資源配置、排程與伺服器運營策略；C-RAN架構如同Amazon Elastic Compute Cloud (Amazon EC2) 或Google的Google Cloud Platform是將資料中心內的資源(主機、網路、儲存設備)視為是一種服務，若前端提出運算需求時，後端系統的工作則是針對所需資源進行配置與排程或啟動新伺服器來執行，本議題以系統角度綜觀分析系統之資源利用率，面對不同經濟價值的用戶，以pay-as-you-go的方式進行收益與成本計算，應用拉格蘭日鬆弛(Lagrangian Relaxation)演算法解題，目標以系統角度在確保服務穩定提供條件下最大化系統的收益。 第5章：動態資源允入與配置策略；計算允入決策的時間、BBU允入組合與配置伺服器開機數量將影響C-RAN利用率，本議題應用背包問題(0/1 knapsack)與裝箱問題(bin-packing)探討在固定決策區間內，以網路角度，分析批次處理BBU的最大價值組合與其裝箱策略，在伺服器容量的條件限制下，以動態規劃法(dynamic programming)與裝箱策略為兩階段演算法，最大化系統收益為目標。 第6章：運算資源轉移最佳化策略；探討系統於實際運營環境中，BBU具有在伺服器間調動和轉移機制下可騰挪出更多資源空間，並因應BBU動態需求變化時系統之整體伺服器運算資源利用率的提升，發展以拉格蘭日鬆弛演算法求得最佳化運算資源轉移方法最大化系統收益。 綜合上述，本論文考量5G行動通訊系統之後端雲端網路C-RAN運算資源管理，著重於將所研發之管理策略應用於未來之5G系統，並且透過計算程式模擬在實際環境與案例參數變化下，歸納所得之最佳管理策略，給予未來5G系統實作之參考並有助於服務品質之提升。
並列摘要
Cloud computing technologies are established with virtualization technologies to form a connected resource processing pool in a fifth-generation (5G) cloud radio access network (C-RAN). This enables operators to reduce capital expenses. However, C-RAN operating expenses are typically ignored owing to the complex challenges of using limited resources to deliver satisfactory quality of service (QoS) to users. This dissertation examines relevant issues of scalability and flexibility in resource management. We proposed a reinforcement learning based system framework to consider an operator’s standpoint to focus on communication (network) and computation (system) perspectives; we analyzed the factors influencing sustainable network evolution, such as task call admission control, resource allocation, scheduling, and migrations in services computing. The problems were formulated as mathematical programming problems. Approaches based on Markov decision processes, dynamic programming, and Lagrangian relaxations were proposed to determine the operating decisions within several practical strategies. These strategies were created to satisfy the QoE requirements of applications and to investigate operating servers within a cost-efficient resource pool. The computational experiment results revealed that the compositions of decisions with task admission, resource allocation, scheduling, and migrations were sufficiently supportive to allow operators to make decisions efficiently and effectively to achieve near-optimal system revenue by leveraging cloud technology in a 5G C-RAN. In some cases, the strategies can serve as valuable references to achieve the optimal solution and some objective values within 3%–10% of the optimal values. The decisions determined from near-optimal solutions are used as guidelines for efficient and effective planning and operations of 5G C-RAN network service providers.