使用OpenFlow階層式統計資料查詢機制以偵測資料中心的大流量

Abstract

在資料中心中如何有效偵測elephant flows (傳輸巨量資料的flow)是一大議題。就網管方面來看，我們能透過偵測這些elephant flows來檢測是否有異常的網路流量；就效能方面而言，為了網路頻寬負載均衡，資料中心目前最好的頻寬負載平衡的方式是Equal Cost Multi-Path (ECMP) routing，但ECMP尋找路由路徑時只考慮到最短路徑，而沒有考慮當時路徑上的頻寬負載情形。相關文獻指出妥善地管理這些elephant flows將能提升113%的整體網路使用率。然而現有可行的方法都有一些缺陷導致它們並不完全適用於資料中心中。有鑑於此，本篇論文善用資料中心裡的“the elephant and mice phenomenon”，基於OpenFlow架構提出了一個階層式查詢的機制。一開始我們會以IP範圍為單位來向OpenFlow交換機查詢特定IP範圍內的flow的統計計數的加總，下一次再將此統計計數總量超出臨界值的IP範圍切割成許多更小的IP範圍並進一步向OpenFlow交換機提出查詢，重覆此步驟直到找到所有的elephant flows為止。另外為了進一步減少此階層式查詢機制，我們又加入了elephant store與IP範圍分裂兩個功能，只需消耗少量網路頻寬便能有效偵測elephant flows。我們使用Mininet軟體定義網路擬真器來進行模擬並且搭配數學期望值分析，結果顯示我們提出的方法能夠有效地減少控制訊息的頻寬消耗以及在更短的時間內偵測出elephant flows。

In datacenters how to effectively detect elephant flows (i.e., flows that transfer significant amount of data) is a major issue. Elephant flow detection can help routing mechanism to redirect elephant flows to the least-loading paths. Equal Cost Multipath (ECMP) is a popular forwarding mechanism to achieve load balancing in datacenter networks. However, ECMP only considers equally distribute the flows over equal-length paths (i.e., the equal number of hops), but does not take the bandwidth loading of each path into account. Previous studies show that to detect and reroute the elephant flows effectively can improve 113% aggregate throughput compared with the simple use of ECMP. A naive way to detect the elephant flows is pulling statistics from each flow independently. However,“the elephant and mice phenomenon” suggests that there are only very few elephant flows in a datacenter network. It’s not efficient to collect each flow’s information. Therefore, we propose a Hierarchical Statistics Pulling (HSP) mechanism using OpenFlow protocol to save the bandwidth consumption by aggregating statistics data. HSP first sends a range query (i.e., aggregate flow statistical request) to obtain the aggregated statistic counters of flows within some IP ranges. If the aggregated byte counter in any of requests is over a threshold, we will further divide the original IP range into smaller ranges and send the range queries of those ranges. These procedures are repeated until elephant flows are detected. In order to further improve the bandwidth consumption of HSP, two supplement functions called elephant store and range splitting are invented. Our approaches can detect elephant flows with only a small amount of network bandwidth consumption. We use Mininet emulator and mathematical expectation analysis to verify our methods. Both of them confirm the benefits of our approaches.