Performance Evaluation of the Advanced NetworkTracker for BitTorrentShakir James, [email protected] (A project report written under theguidance of Prof. Raj Jain)DownloadAbstract:BitTorrent, a popular Peer-to-Peer (P2P) protocol, has recently engendered considerable controversy. ManyP2P applications implement a variant of the protocol due to its cost-effective scalability. However, forInternet Service Providers (ISPs) the protocol is notorious for its high network cost. To reduce these costs,ISPs have deployed network devices to detect and throttle BitTorrent traffic. In response, applicationdevelopers have begun encrypting the protocol's traffic to avoid ISP-detection. These actions perpetuate a"cat and mouse" game between ISPs and developers. Recent proposals to reduce BitTorrent's network costinvolve either changing the protocol or deploying ISP-managed services. These proposals require broadadoption by developers to significantly benefit ISPs. In this paper, we propose a solution that is universallyapplicable to all BitTorrent clients and provides savings for ISPs out of the box.We have developed a high performance network device, called the Advanced Network Tracker (ANT),which reduces BitTorrent network cost without changing the protocol. Moreover, ANT requires no explicitcommunication between client applications and ISPs. It inspects packets on the wire to identify BitTorrenttraffic to discover local peers sharing similar files. To reduce network costs, ANT localizes traffic to an ISP'snetwork by informing local peers about each other. In this work, we implement a working prototype of ANTusing Network Processors (NP) for high throughput packet processing. In addition, we evaluate ourimplementation on the Open Network Lab's [ONL] NP-based programmable routers. Our evaluation showsANT has a considerable impact on cross-ISP traffic. And equally important, it has a negligible impact onclient performance.Keywords:Network architecture, network processors, traffic management, performance evaluation, peer-to-peer (P2P),BitTorrentTable of Contents1. Introduction2. Background3. Design3.1 Component Architecture3.2 Packet Filter3.3 Local Tracker3.4 Proxy CachePerformance Evaluation of the Advanced Network Tracker for BitTorrent1 of 154. Implementation4.1 Network Processors4.2 Open Network Lab4.3 ONL Network Processor Router4.4 ANT Plugin5. Evaluation5.1 Goal and Metrics5.2 Factors5.3 Evaluation technique and Experimental Design5.4 Results6. Related Work7. Future Work8. Concluding RemarksAcknowledgementsReferencesList of Acronyms1. IntroductionPeer-to-Peer (P2P) systems promise cost-effective file distribution for content providers on the Internet. In aP2P system, each peer downloads as well as uploads data to other peers. Thus, P2P systems are self-scalingbecause their bandwidth capacity increases as the number of peers increase. Compared to the client-servermodel or content distribution networks, P2P systems are cheaper for content providers since they do notrequire an investment in dedicated servers. Furthermore, P2P systems are better able to recover from failuressince each peer holds some piece of the file. This paper focuses on the BitTorrent P2P protocol. Researchershave shown that the BitTorrent protocol is self-scaling [Yang04][Qiu04], robust [Legout06], and providesnear optimal performance in terms of peer download time [Bharambe06].Although the BitTorrent protocol provides cheap scalability for content providers, it is expensive for InternetService Providers (ISPs). The protocol is oblivious to the underlying network topology, so its traffic traverseroutes without concern for ISP economics. In fact, the protocol increases traffic on costly, cross-ISP link[Bindal06]. At the same time, P2P applications like BitTorrent increase revenue for ISPs because theyencourage more broadband subscriptions. As a result, ISPs are left in a quandary: throttle BitTorrent traffic toreduce network cost at the risk of losing revenue from broadband subscriptions. ISPs have deployed trafficshaping devices [P-Cube][Sandvine] to "throttle" or limit BitTorrent traffic. Although the legality of trafficshaping is debatable, the practice degrades P2P application performance and can turn away potentialcustomers.Instead of reducing BitTorrent network cost, traffic-shaping devices have sparked an arms race betweenBitTorrent developers and ISPs. To evade identification by the first generation of these devices [P-Cube],some BitTorrent applications added header encryption to peer-to-peer messages. To counter, ISPs introducednew devices [Sandvine] that targeted other protocol (peer-to-tracker) messages. Recently there have beensigns of an armistice, as ISPs have begun supporting initiatives to cooperate with developers to reduceBitTorrent network cost [Xie08]. Some authors [Bindal06][Xie08][Choffnes08] have proposed changes to thede facto protocol. Unfortunately, these collaborative initiatives have not been widely adopted.As a new solution, we propose a network device, called the Advanced Network Tracker (ANT), which ISPsinstall on their network beside edge routers. What separates ANT from other proposals, is that ANT workswith all BitTorrent clients and requires neither protocol changes nor direct application-ISP communication.Thus, ANT's novel contribution is solving ISPs' P2P quandary without requiring any extensions to thePerformance Evaluation of the Advanced Network Tracker for BitTorrent2 of 15protocol. To reduce network costs, ANT keeps BitTorrent traffic within an ISP's local network.We have implemented ANT on Intel's Internet eXchange Architecture (IXA), which uses a NetworkProcessor (NP) called the IXP [IXP] for high performance and programmability. NPs are a programmablealternative to Application Specific Integrated Circuits (ASIC) for rapid development of high performancenetwork devices. We used the specialized hardware units on the IXP, such as the hash unit, a TernaryContent ddressable Memory (TCAM), and high performance memory interfaces to enable high performancepacket processing. We also used the hardware multithreading capability of the IXP's Micro-Engines (ME) tohide the latency of accessing of chip memory. To evaluate our implementation, we deployed ANT on theOpen Network Lab's [ONL] IXP-based routers. Our evaluation shows that ANT has a meaningful impact oncross-ISP link utilization and an insignificant impact on client download time.The remainder of the paper is organized as follows.
View Full Document
Unlocking...