Pricing and Peering Strategies of Differentiated Services Content Networks ¨ Ozg¨ ur Erc¸etin * , † Sabanci University Faculty of Engineering and Natural Sciences Istanbul 34956, TURKEY. oercetin@sabanciuniv.edu Leandros Tassiulas Institute for Systems Research University of Maryland College Park, MD 20740, USA. leandros@isr.umd.edu Abstract Web sites disseminate some of their information to surro- gate caches in order to reduce the latency observed dur- ing the delivery of the information. The surrogates clas- sify publishers under several classes with respect to their willingness-to-pay. Surrogate partitions the total cache ca- pacity among different classes to provide a loose version of Quality of Service. In our model, publishers try to get as large cache space as possible, while the surrogate is required to achieve fair allocation among the publishers. Specifically, each publisher should be charged the same if they receive equal share of caching space. We determine the optimal pricing strategy of the surrogate maximizing its revenue. We also analyzed the competition between two sur- rogates under this model and determined the condition that leads to a Nash equilibrium. We showed that at equilibrium surrogates peer with each other as if there is a single com- bined surrogate server. 1 Introduction As the Internet becomes more commercially oriented each passing day, the effects of the user latency is becoming more prominent. In the new Internet economy, the web sites (origin servers) are publishers of Internet content. They ap- pear as businesses generating their revenues from the con- tent that they provide. Usually, the more popular a web site is more revenue it generates. Meanwhile, the popular- ity of a web site depends among other things on the fast delivery of its content. The web sites began using surro- gate caches to deliver their content with low latency to the users. However, until recently the routing of the user re- quests as well as the dissemination of content from the ori- * Corresponding Author. † This work was in part done, when the author was with the Department of Electrical and Computer Engineering, University of Maryland College Park. gin servers to these proxies was done in a best-effort fash- ion. Several Content Distribution Networks (CDNs) such as Akamai, Edgix and DigitalIsland began offering caching services to the publishers. CDNs provide the publishers a reliable and robust surrogate caching server infrastructure, which covers most of the globe. The origin servers ‘rent’ portions of the caching capacity from the CDN’s caching servers and receive varying benefits depending on the sur- rogates’ locations, capacities and user request arrival rates. The provision of Quality of Service (QoS) will be an im- portant and integral property of the future network commu- nications. Current research efforts in network QoS identi- fied the provision of statistical QoS guarantees as the more realistic approach as compared to the provision of hard guarantees. The diffserv architecture is a model developed under this observation. The diffserv architecture offers lim- ited service guarantees in the form of service classes [1]. The Olympic service model was proposed as an example of diffserv architecture [2, 3]. In this model bronze, silver and gold service classes are offered to the users. The packets assigned to the gold service class usually experience lighter load than the packets assigned to the silver or bronze class. However, there are no guarantees for the service quality. Thus, at certain times load for a service class may be so high that the performance of an application run with gold class can be less than satisfactory. Meanwhile, at other times the total network load may be so low that the same application can be satisfactorily run even with bronze class. In the following we will explore the provision of simi- lar kind of statistical QoS guarantees in Content Delivery Networks. We assume that each surrogate offers a limited number of service classes to the publishers. The publishers have different smooth and concave willingness-to-pay (util- ity) functions for the resource. For simplicity the CDN as- signs each publisher into one of few available classes. The surrogates reserve fixed amount of resources for each ser- vice class. The publishers that subscribe to the same class share the same resource. Thus, each publisher is assigned a caching space depending on the service class and the num-