一個端對端的壅塞控制機制的設計與效能分析

Abstract

廣泛且簡易的網路應用造就網際網路的成功，而TCP 是目前網際網路上最被廣為使用的傳輸層協定。這是因為TCP 利用其可靠性的傳輸以及端點對端點的擁塞控制之特點，為傳統的網路環境中，提供了一個可接受的服務品質。無論在學術界還是工業界，TCP/IP 的研究始終是一個熱門的課題。 隨著蓬勃發展的網路硬體技術與網路應用，可預見的是，使用者的網路頻寬會越來越大與無線環境也會越來越普遍。這將會造成網際網路的訊務流量快速成長以及產生多樣化的傳輸方式，再加上使用者對於網路效能的要求也會不斷增加。面對此一改變，如何有效的使用網路資源、適應網際網路寬頻化與異質化網路環境(有線、無線網路的混接)以及滿足使用者的要求是一個成功的壅塞控制機制所要面臨的種種問題。 另外，假若TCP 的改良無法齊頭並進，那麼很快的，TCP 本身的壅塞控制機制將成為效能上的瓶頸。至今TCP 被創造出許多不同的版本，用來改進網路的使用效能。這些版本可以分為兩大類：loss-based TCP 和delay-based TCP。雖然很多研究報告已經指出delay-based TCP 的機制可以做出較精確的控制、擁有更好的特性， 使得delay-based TCP 在很多方面都要優於loss-based TCP，例如：整體網路的使用率、穩定性、公平性以及傳輸速率…等。但是loss-based TCP 仍為現今主流，而且目前關於TCP 的研究大多集中在loss-based TCP 上，原因在於這類TCP 的方法實做起來較簡單且競爭力較另一類TCP 來得要強，故TCP SACK 方案為現今網際網路上最被廣為使用的TCP 版本。很可惜的是，實際上這兩類TCP 都仍然有一些屬於自己的缺點存在於它的壅塞控制機制中，因此衍生出一個問題：要如何設計TCP 以減少本身的缺點。 在這份論文中，我們提出一個新的端對端的壅塞控制機制用於適應新的網路環境、充分利用網路資源以滿足使用者的要求以及克服TCP 本身缺點所造成的負面影響，並命名為Medley TCP。與傳統的TCP 相比較，最大的不同在於我們重新設計TCP 整體的壅塞控制機制與架構。具體來說，Medley TCP 是以loss-based TCP 的演算法做為基礎模組，搭配delay-based TCP 的估算機制與控制模式來調整最終的傳輸速率，嘗試結合這兩類TCP 的優點與精神。為了讓連線在高頻寬-延遲乘積網路中能有較好的表現，Medley TCP 利用連線估測堆積在佇列中的封包數量與其推算之值為依據，使得一個連線的送端在調整壅塞窗口大小時採取更有效和積極的態度。另外，Medley TCP 本身的特性可以準確判斷出封包遺失是屬於網路的壅塞還是傳輸失誤所造成，透過封包遺失原因的分類，Medley TCP 可以適切的對不同原因的封包遺失做出不同的反應，因此改善了在異質化網路環境中TCP 傳輸的效能。 通過廣泛網路模擬軟體的實驗測試和網際網路的實地測試，Medley TCP 被證實能在高頻寬-延遲乘積網路與異質化的網路環境下獲得大幅度的效能提高，而且Medley TCP 也保留了delay-based TCP在整體網路的使用率、穩定性、公平性以及傳輸速率…等方面的優異特性。更重要的是Medley TCP 可以與現存的TCP SACK 和諧共存而不降低其效能，這是因為Medley TCP 可以更有效的利用網際網路以獲得頻寬，而不是搶奪其他連線所應佔用的資源。最後，Medley TCP不用修改連線的接受端或者中間節點，僅需要在資料的發送端做修改即可。因此，Medley TCP 可以容易地安裝於實際的網際網路中。

The success of the Internet can be attributed to the large number of useful applications which are easily executed by a user for running on the Internet. Transmission Control Protocol (TCP) is a widely deployed end-to-end transport protocol in the current Internet. This is because TCP provides an acceptable service with a reliable data transport as well as controls the connection's bandwidth usage to avoid network congestion for two end hosts in the Internet. Nowadays, the research of TCP/IP is still a hot topic in both the academia and the industry. With the fast growth of Internet hardware, technologies, and applications, the network bandwidth of a user is getting higher and wireless links are more and more popular everywhere. This will make Internet traffic increase quickly and the ways of data transport vary. Also, users' demands for network performance are getting stricter. Facing these challenges, how to efficiently utilize network resources, how to work well in both high bandwidth network and heterogeneous network (mixed with wired and wireless networks), and how to satisfy users' requests are essential issues to a successful congestion control mechanism. In addition, if the modification of TCP's congestion control does not catch up the Internet change, the performance bottleneck will soon be TCP itself. Up to now, in order to improve network utilization, TCP has several implementation versions which can be classified into two categories, loss-based TCPs and delay-based TCPs. Although it has been demonstrated that the delay-based TCP outperforms the loss-based TCP in the aspects of overall network utilization, stability, fairness, throughput, etc., in the real Internet, loss-based TCPs are still the mainstream and remained as the dominant algorithm used in practice. Furthermore, the implementation of a loss-based TCP is easier than that of a delay-based TCP, and a loss-based TCP could get more resources than a delay-based TCP could when they coexist in the same network, so many studies focus on loss-based TCP algorithms, and one version of loss-based TCP mechanisms, TCP SACK, has been widely deployed on the Internet. However, in fact, loss-based and delay-based TCPs have their own advantages and shortcomings. Hence, there is an issue ``how do we design the algorithm and architecture of TCP?'' In this dissertation, we propose a novel end-to-end congestion control mechanism called Medley TCP that is able to efficiently utilize Internet resources, adapt itself to a new network circumstance, satisfy users' requests, and accommodate shortcomings of the conventional TCP. Medley TCP differs from the traditional TCP extremely in that we redesign the algorithm and architecture of the whole congestion control mechanism of TCP. Specifically, Medley TCP tries to combine advantages and characteristics of both loss-based and delay-based TCPs, and therefore incorporates a scalable delay-based component into a loss-based TCP algorithm. This scalable delay-based component has a rapid window increase rule when the network is sensed to be under-utilized and gracefully reduces the sending rate once the bottleneck queue is built. Therefore, Medley TCP connections will react faster and better to high BDP networks and improve the overall performance. Moreover, we utilize the innate nature of Medley TCP to detect congestion losses or random packet losses precisely. Through the packet loss differentiation, Medley TCP reacts appropriately to the losses and consequently the throughput of connection over heterogeneous networks can be significantly improved. Extensive experiments of a network simulator and real world Internet traffic measurements have been conducted and show that Medley TCP not only can improve the performance significantly over high BDP networks and heterogeneous networks but also can keep the good characteristics of delay-based TCPs in the aspects of overall network utilization, stability, fairness, throughput, and so on. Importantly, Medley TCP does not cause any detrimental effects to TCP SACK, and vice versa when they share same resources in the networks: that is, Medley TCP connections achieve higher throughput by using bandwidth that will not be used by SACK connections anyway, not ``stealing'' bandwidth from SACK's connections. Finally, Medley TCP only involves modification at the sender without requiring changes of the receiver protocol stack or intermediate router nodes. As a result, it can be easily deployed in the current Internet.