具補償機制的壅塞感知負載平衡系統

Abstract

在資料中心網路通常會使用多根拓墣，能夠低成本的提供多條路徑，使點與點之間的頻寬增加。因有多條路徑，必須善用所有的可用路徑，並避免走壅塞路徑，故需要有良好的負載平衡機制來做路徑選擇。 目前常見的負載平衡機制都只考慮最短路徑集，在最短路徑集中選擇不壅塞的路徑或是盡可能將封包平均分配在各路徑。這些方法僅能在對稱的拓墣有良好的效果，一但出現複數個連線異常，會造成某些主機之間的最短路徑減少，進而造成某些主機之間不必要的頻寬縮減。 本論文提出了一個具有補償機制的負載平衡機制，能夠在網路出現複數個連線異常時，啟動補償機制。透過選擇最短路徑與非最短路徑，來補償因連線異常所造成的頻寬損失，進而增加點與點之間的頻寬與縮短流量傳輸完成時間。

In order to provide large bisection bandwidth, data center networks employ multi-rooted topologies (e.g., Leaf Spine, Fat Tree). There are multiple paths can be used between any two hosts. Therefore, data center needs a good load balancing mechanism to utilize the available bandwidth. Recent state-of-art load balancing mechanisms only load balance traffic among all the shortest paths. These mechanisms only perform effectively when the topology is symmetric. If there are link failures occurred in the network, the number of shortest paths diminished. Using these mechanisms might cause unnecessary bandwidth reduced. This thesis presents CACOM, a congestion-aware load balancing with compensatory mechanism. The compensatory mechanism will be activated when there are link failures in the network. This mechanism increases the bandwidth by using both shortest path and non-shortest path. The compensatory mechanism can increase the bandwidth between switches and reduce the flow completion time of the flows in the network.