適用於軟體定義網路之動態自適應負載平衡演算法

Abstract

中文摘要 近年來，網際網路的使用者已戲劇化的成長。同時，雲端服務如雲端儲存、雲端計算的需求也緊接著成長。另一方面，大多數在資料中心執行的應用程式，需要更多的頻寬來提供更高的效能。 資料中心網路常採用能以低成本提供多條路徑的多根拓樸。等價多路徑 (ECMP) 是一個被廣泛採用的動態路由演算法，能夠利用路徑的多樣性，提供節點負載平衡。然而，等價多路徑演算法應用於負載平衡乃是一個靜態演算法，並沒有考量傳輸埠的利用率。該演算法可能會利用封包開頭產生的雜湊值可能會碰撞，以至於網路有堵塞的現象。 本論文提出了動態適應負載平衡演算法 (DALB algorithm)，利用軟體定義網路控制器的知識，監控並且偵測網路賭塞的情形。由於軟體定義網路的控制器，具有整個網路的全局關，透過蒐集網路統計數據，軟體定義網路的控制器能夠取得目前網路的附載，以及各傳輸埠的利用率。當有連線被高度利用十，動態適應負仔演算法，將會把流量進行新的路徑規劃，重新導向另一個較不壅塞的路徑。 我們的研究結果顯示，動態適應負載演算法能夠提高更高的頻寬，迴避應用程式碰到網路壅塞的問題，並可以更佳的平衡整體網路的附載。同時，動態適應負載演算法消耗的時間相當短，封包可以在微秒單位下傳遞，且能偵測高負載與重導向流量。實驗數據指出，在即時重導向流量的情境下，並沒有封包遺失的情形，顯示動態適應負載演算法具有相當可靠。

Over the past few years, the number of internet users has increased dramatically. Also, the demand for cloud services such as cloud storage and cloud computing are increased as sequentially. On the other hand, most of the applications that are running in the data centers require more bandwidth to provide higher performance. Data center networks (DCN) adopted multi-rooted topologies which provide multiple paths with low costs. Equal Cost Multipath (ECMP) is widely used in DCN as a dynamic routing algorithm that takes advantage of path diversity. It also balances the load among the node ports. However, ECMP is a static load balancing with no means to take the utilization on the outgoing ports into account. Moreover, it is based on hashing the packet’s header which could produce collision and can lead to a traffic congestion. In this work, we proposed the Dynamic and Adaptive Load-balancing (DALB) algorithm, which uses the SDN controller knowledge for monitoring and detecting congestion. SDN controllers have a global view of the whole networks. From the gathering of network statistics, SDN controllers can obtain the current load in the links and ports of the entire network. When the link is highly loaded, the DALB algorithm routing function redirects the traffic flows to less congested one. Our results have shown that DALB provides higher bandwidth to applications reduces congestion problems and is capable of balancing the traffic loads. Also, DALB takes very short time, (i.e. Packets can reach destinations in milliseconds) and from detecting high load to the actual rerouting of flows. The experimental results are indicative of DALB’s reliability in ensuring no packet loss while rerouting of live traffic streams.