An Analysis of Internet Content Delivery Systems Himani Apte CS 739 Distributed Systems University of Wisconsin Madison January 20 2006 Spring 2006 1 Overview of the Internet The TCP flows were reconstructed at the monitoring hosts to extract information to categorize them into HTTP and non HTTP traffic The HTTP traffic is further distinguished into WWW Kazaa and Gnutella based on the destination ports Akamai traffic is identified based on whether it is served by an Akamai server This methodology misses the internal traffic within the local network of the University for instance file sharing traffic among Kazaa users within the University network The analysis also does not take into account the nonHTTP TCP traffic that is 43 of the total TCP traffic and non TCP traffic that is 3 of the total network traffic Other peer to peer systems such as BitTorrent and Napster have also been excluded from the study The paper analyzes four different content delivery systems from the arena of client server oriented world wide web content delivery networks and peer to peer file sharing systems The important features of peer to peer systems are symmetry among the peers since they behave as both servers and clients and scalability up to as many as millions of machines Dynamic membership wide area network and heterogeneity of the participating systems in terms of their bandwidth connectivity and performance are some of the distinguishing characteristics of peer topeer systems Such systems tend to be application specific Some peer to peer systems may have a hierarchy among the members for instance some peers in Kazaa are supernodes and maintain indexes for the content available at peers in the nearby neighborhood 4 Observations This section summarizes the observations made in class about the analysis presented in the paper 2 Problem Statement The paper examined the traffic flow of content delivery 4 1 Data characteristics systems focusing largely on web versus peer to peer traffic flows and specifically on HTTP web traffic Akamai The HTTP trace summary statistics presented in Table 1 are unavailable for outbound Akamai traffic as there Kazaa and Gnutella delivery systems are no Akamai servers hosted within UW Kazaa has the highest outbound traffic in terms of net bytes transferred in spite of a much smaller server and client population 3 Methodology within UW The methodology employed was passive network moniThe total TCP bandwidth consumed by HTTP transfers toring of all traffic coming in and out of the border routers for different content delivery systems presented in Figure between the University of Washington UW and the rest 1 show a typical diurnal pattern WWW traffic peaks in 1 is actually a measure of the proportion of unique bytes accessed to the net bytes accessed The ideal byte hit rate Figure 14 for outbound Kazaa traffic was found to stabilize at 85 while that for inbound traffic did not stabilize by the end of the trace The cache byte hit rate as a function of population size is presented in Figure 15 On the one hand increasing the number of Kazaa clients may increase the number of requests thereby lowering the cache hit rate On the other hand this leads to the complementary effect of caching in the numerous clients thereby improving the cache hit rate Although the preliminary investigation presented in the paper suggests that caching would have a large effect on a wide scale P2P system potentially reducing wide area bandwidth demands dramatically it may not actually be 4 2 Content delivery characteristics feasible to employ caches in P2P systems due to legal isMost bytes are transferred in video objects although most sues pertaining to the content distributed in such a system requests are for GIF and JPEG images The median object Also the paper does not give an insight into a realistic size size for WWW is 2 KB while that of peer to peer systems of cache that would be sufficient to obtain improvements is 4 MB in bandwidth usage The top bandwidth consuming UW clients Figure 7 and UW servers Figure 10 are the Kazaa peers Hence caching will be most beneficial for Kazaa file sharing 5 Conclusions system The cause for large bandwidth consumption are large size objects and very popular objects that result in The paper presents a quantification of the domination of large number of requests As a result small number of P2P systems in the modern day Internet traffic Although the global characteristics are not easily seen clients consume a large amount of bandwidth in peer tolooking at a small part of the network UW in this case it peer systems still makes interesting revelations about the network trafThe Kazaa and Gnutella servers are not perfectly loadfic flows In the future we would expect WWW traffic to balanced in spite of the scalability of peer to peer sysshow similar characteristics and an even larger P2P traftems Possible cause for this may be the existence of fic highly popular content on a single server or availability of large size objects on only a few peers However it is incorrect to make conclusions about whether or not the Kazaa and Gnutella servers are load balanced based on the data available as the trace does not include internal file sharing traffic within UW network true daylight hour as opposed to Kazaa traffic which peaks late at night The HTTP trace was collected in May June over a nine day period The trace could significantly vary based on the timing when it was collected as the network behavior of university students may be widely different during summer break than in the final exams week Also a longer data sample would be desirable Analysis of the UW client and server TCP bandwidth presented in Figure 2 indicates that Kazaa peers within UW act as servers much more than the web servers at the university A possible reason could be high connectivity of the UW Kazaa peers 4 3 Role of caching For studying the role of caching in CDNs and P2P systems the authors have simulated infinite capacity caches The three causes of cache misses popularly known as the three C s are cold capacity and conflict misses In case of infinitely large cache the only kind of misses that can occur are cold misses As a result the cache miss rate 2