針對資料中心之不需要大流量偵測的低延遲多路徑繞徑

Abstract

中文摘要

現代的資料中心提供了巨量的頻寬給數以萬計的機器。然而，即使因為在最上端的交換器的埠數限制，資料中心的拓樸結構通常對於任一組host之間仍存在擁有相同花費的路徑的multi-rooted tree。現有的多路徑方案通常都仰賴每一個flow的靜態的雜湊，而並沒有去區別長的flow以及短的flow。但是因為發生在長的flow中hash collision的問題，導致它無法完全利用可用的頻寬。而且它也會在短的flow的egress port buffer產生head-of-line blocking問題。多路徑繞徑是一個新的受歡迎的技術，它保護資料中心的網路免於因為尖峰負荷或是連結故障所造成的突然的壅塞。然而，多路徑路由協議需要將流量分散跨越至多條路徑上，以達到最佳的頻寬使用率。目前的分切方案在切割的精細度以及可用的頻寬使用率之間做取捨。在封包的層級進行流量切割可以快速又準確的分配流量給每一條路徑，但是卻可以在TCP flow中重新排列封包，導致混淆TCP壅塞控制。另一方面，在flow層級進行流量切割可以容易地解決封包重新排序以支援TCP壅塞控制，但是卻致使在長的flow之間發生的hash collision，而導致不良的可用頻寬使用率以及增加全部的flow的完成時間。在這篇論文中，我們提出一個簡單而有效率的不需要大流量偵測的多路徑繞徑方法。我們使用OpenFlow交換器去實作一個以VLAN為基礎的繞徑來說明成本效益以及擴充性的挑戰，以及一個以hard time-out為基礎的flow移除功能去將長的flow分切成短的flow並且分配他們到所有可能的路徑去，以達到一個在資料中心網路的有效率且動態的流量負載平衡。根據使用mininet的模擬結果，我們對於整體的flow的完成時間減少了44%，以及降低了整體的交換器的 flow table entry的消耗達32%。

關鍵字：多路徑繞徑、OpenFlow、Time-out、大流量、VLAN、資料中心

Modern data centers offer huge aggregate bandwidth to clusters of tens of thousands of machines. However, because of limited port densities even in the highest-end switches, data center topologies typically consist of multi-rooted trees with many equal-cost paths between any given pair of hosts. Existing multipathing schemes usually rely on per-flow static hashing and does not differentiate between elephant and mice flows. As a result it does not fully utilize the available bandwidth due to hash collision among elephant flows. Further it also creates head-of-line blocking for mice flows in the egress port buffer. Multipath routing is a popular recent technique that protects data center networks from sudden congestion caused by load spikes or link failures. Multipath routing protocols however, require schemes for splitting traffic across multiple paths to achieve optimum bandwidth utilization. Current splitting schemes present a tussle between slicing granularity and available bandwidth utilization. Splitting traffic at the granularity of packets quickly and accurately assigns the desired traffic share to each path, but can reorder packets within a TCP flow, confusing TCP congestion control. On the other hand splitting traffic at the granularity of flows can readily resolve the packet reordering to support TCP congestion control but leads to hash collision among elephant flows, resulting poor available bandwidth utilization and increases overall flow completion time (FCT). In this work, we propose a simple yet effective multipath routing scheme that does not need elephant detection and uses the OpenFlow switches to implement a VLAN based routing to addresses the challenges of cost-effective scalability and a hard time-out based flow removal feature to break down elephant flows into mice flows and distribute them among all possible paths to achieve an efficient and dynamic traffic load balancing in data center networks. As per our simulation using mininet, we found a 44 % reduction in overall FCT and 32 % reduction in the consumption of overall flow table entry resources.