雲端資料中心之動態且互斥之資料鏈結層繞徑研究

Abstract

近年來，隨著雲端運算 (Cloud Computing) 的快速發展，資料中心的網路流量也越來越繁重。在這篇論文中，我們討論資料鏈結層(Data Link Layer, Layer-2)中的繞徑議題，稱作「動態且互斥之節點分離生成樹繞徑法」(Dynamic and Disjoint Edge Node Divided Spanning Tree, D^2 ENDIST)，用來改善資料中心的網路流量分配不均的狀況。 在本文，我們提出的「動態且互斥之節點分離生成樹繞徑法」由兩個主要的概念所構成：(1) 靜態的互斥的節點分離生成樹 (Disjoint Edge Node Divided Spanning Tree, DENDIST) 及 (2) 動態規劃路徑 (Reroute by Dynamic Reweights)，前者計算出多重路徑，且路徑之間是不相重疊或最少重疊的以減少流量壅塞；後者則是動態的平衡流量。本文也利用一系列的模擬數據來驗證我們提出的正確性及優越性。

This thesis presents an improved layer-2 routing algorithm, called dynamic and dis-joint edge node divided spanning tree (D2ENDIST), to overcome the issues of the single path route and unbalanced link utilization in cloud datacenters. D2ENDIST consists of two key schemes: (1) disjoint ENDIST routing and (2) reroute by dynamic reweights. The former scheme can provide multi-path routes, thereby reducing traffic congestion. The latter scheme can balance the traffic load and improve the link utilization. Our experimental re-sults show that the proposed scheme can enhance system throughput by 25% subject to the constraint of very short failure recovery time compared to the existing ENDIST scheme.