基於軟體定義網路之資料中心動態負載平衡路徑最佳化機制

Abstract

近年來，在資料中心內交換器之間傳送的資料流量愈來愈多，此容易導致交換器之間的路徑擁塞，因而導致延遲時間變長及產量降低。軟體定義網路(SDN)架構將交換器資料層與控制層分離，並將流量控制功能集中到控制層的控制器上，如此可以讓網路管理者能夠透過外在的控制器有效地管理資料中心內部網路中的資料流。然而，現有基於SDN之路徑最佳化演算法並無法一體適用於不同的資料中心網路拓樸。本論文的主要貢獻為：(1)提出一個具網路拓樸通用性的資料流路徑動態變更機制(DLPO)。透過軟體定義網路之特性，管理資料中心內部網路中所有的資料流流向，在封包傳遞期間視需要動態改變資料流之路徑，以有效因應不同的資料中心網路拓樸，解決資料流因為網路壅塞而造成的封包延遲問題。(2)我們也提出了一個基於優先度之流量表更新機制，確保在新路徑上各交換器的流量表都完全被更新之後，資料流才立即被導至新路徑，如此可以有效避免因為路徑改變而造成封包遺失及資料中心網路產量下降的問題。我們所提的方法由兩個演算法組成，分別是能快速平衡資料中心內部網路負載並解決大部份路徑壅塞的multi-link DLPO演算法與能有效解決剩餘路徑壅塞的single-link DLPO演算法。實驗結果證明，在hot-spot模式的資料流下，我們所提出之具網路拓樸通用性之資料流路徑動態變更演算法(DLPO)，相對於現有具代表性資料流路徑動態變更演算法(LABERIO)，在fully populated拓樸中能提升24.8%的產量與8.3%的頻寬利用率，在fat-tree拓樸中則能提升18.9%的產量與9.5%的頻寬利用率。此外，在fully populated拓樸中，當流量負載由0.1持續增加至0.9時，DLPO演算法能提升25.6%的平均產量。這證明了DLPO演算法在資料中心中的可擴展性。

Nowadays, more and more data transmitted in data centers may lead to the congestion problem between switches and result in long delay and low throughput. Software-Defined Networking (SDN), which decouples the data plane and control plane, allows a network administrator to perform flow management in a network with an external SDN controller. Related work tried to optimize paths assigned to flows dynamically to improve performance of SDN-based data center networks. However, existing SDN-based path optimization algorithms cannot be applicable to different data center network topologies. The main contributions of this paper are as follows. (1) We propose a Dynamic Load-balanced Path Optimization (DLPO) algorithm which can suit for different SDN-based data center network topologies. The proposed DLPO change paths of flows during flow transmissions, achieve load balancing among different links, and efficiently resolve the network congestion problem in a SDN-based data center network. (2) We also propose a priority-based flow table updating strategy to ensure that flows of a congested path will be redirected to a light-loaded path as soon as all flow tables of the associated switches in the light-loaded path have been updated successfully so as to avoid packet loss caused by changing paths of flows. The proposed DLPO is composed of two algorithms, which are multi-link DLPO algorithm and single-link DLPO algorithm. The multi-link DLPO algorithm can balance the link loads in a network quickly to resolve some congested paths and the single-link DLPO algorithm can reroute flows to avoid using links with large loads to resolve congested paths that the multi-link DLPO algorithm cannot handle. Simulation results show that the proposed DLPO increases 24.8% and 8.3% of throughput per flow and bandwidth utilization compared to LABERIO in the fully populated topology, respectively under a hot-spot traffic pattern. In addition, DLPO increases 18.9% and 9.5% of throughput per flow and bandwidth utilization compared to LABERIO in the fat-tree topology, respectively under a hot-spot traffic pattern. Simulation results also show that the average throughput per flow of the proposed DLPO is 25.6% higher than that of LABERIO as traffic load continually increases from 0.1 to 0.9 in a fully populated topology. This shows that the proposed DLPO is scalable for SDN-based data center networks.