在異質IP網路中利用壅塞遺失比提升TCP效能之研究

Abstract

TCP協定是許多重要的網路應用如網頁瀏覽、檔案下載等軟體所使用的傳輸工具。但不幸的是，當應用TCP於無線網路時，效能會嚴重降低。傳統TCP假設所有封包遺失(packet loss)皆是因為網路壅塞(Congestion Loss)而遭路由器丟棄(router drop)所致。然而當無線網路介面成為網路傳輸路徑之一部分時，封包還有可能因訊號突然受雜訊干擾而使封包傳送失敗造成遺失(Random Loss)。 在這篇論文中，我們首先提出了「壅塞遺失比」(Congestion Loss Ratio; CLR)的觀念及其估計方法。「壅塞遺失比」(CLR)指的是在所有最近遺失的封包中，真正因網路壅塞而遭路由器丟棄的封包數之比例的估計值。CLR值的估算主要來自於記錄大小封包在行經有線與無線網路路徑時，遺失封包個數的比例而得出。 我們接著利用CLR觀念，提出了FACK+與TFRC+兩種TCP流量控制方法並利用ns-2網路模擬軟體進行模擬分析。模擬結果顯示：(1) CLR觀念可適用於路由器上常見的Drop-Tail與RED兩種流量管制(Traffic Management)方法。（2）FACK+與TFRC+的確能有效提升TCP的效能。舉例來說，當無線網路介面上每位元組平均資料錯誤率(BER; Byte Error Rate)為0.005%、平均網路背景流量為5%時，FACK+與TFRC+相對於其原始設計，分別能有效提升40%與220%的效能。

The Transmission Control Protocol (TCP) is the transport protocol for many important applications such as web browsing and file transfer. Unfortunately, when TCP is used over wireless networks, it usually suffers severe performance degradation. The traditional assumption for TCP, that all packet losses are due to congestion losses at intermediate routers, is no longer valid. When wireless networks become a part of the network transmission path, chances are packets might be corrupted due to interference or noise which creates random losses. In this thesis, we proposed the idea of Congestion Loss Ratio (CLR). CLR refers to the estimated ratio of the number of congestion losses to that of all packet losses. CLR is derived from observing how different-sized packets are lost on wired and wireless networks. We then propose two TCP congestion control algorithms, FACK+ and TFRC+, both based on CLR, and then use the network simulator ns-2 to evaluate the performance of our proposed algorithms. Simulation results show that: (1) CLR could be applied to conventional router-traffic-management mechanisms, such as Drop-Tail and RED.（2）FACK+ and TFRC+ effectively improve the performance of TCP. For example, when the average BER (Byte Error Rate) on wireless network is 0.005% and the average background traffic is 5% of the bottleneck link bandwidth, FACK+ and TFRC+, relative to the baseline TCP algorithms, improves TCP performance by 40% and 220%, respectively.