Modeling content delivery networks and their performance
Introduction
Few things compare with the growth of the Internet over recent years. A key challenge for Internet infrastructure has been delivering increasingly complex data of different types and origin to a growing user population. The need to scale has led to the development of clusters [1], global content delivery networks (CDN) [2] and, more recently, P2P structures [3]. However, the architecture of these systems differs significantly, and the differences affect their performance, workloads, and the role that caching can play [4], [5].
CDNs are overlay networks across the wide-area Internet which consist of dedicated collections of servers, called surrogates, distributed strategically throughout the Internet. The main aim of the surrogates is to be close to users and provide them with content in a low-latency mode. The surrogates are normally proxy caches that serve cached content directly with a certain hit ratio; the uncached content is previously obtained (if possible) from the origin server before responding. When a client makes a request for content inside a CDN, it is directed to an optimal surrogate, which serves this content within low response time boundaries—at least compared to contacting the origin site [6], [7]. CDNs such as Akamai [8] or Digital Island [9] are nowadays used by many websites as they effectively reduce the client-perceived latency and balance load [10]. They accomplish this by serving content from a dedicated, distributed infrastructure located around the world and close to clients. The content is replicated either on-demand, when users request it, or replicated beforehand, by pushing the content on the content servers [11], [12]. CDN services can improve client access to specialized content by assisting in four basic areas:
- •
Speed, reducing the response and download times of site objects (e.g. streaming media), by delivering content close to end users.
- •
Reliability, by delivering content from multiple locations; a fault-tolerant network with load balancing mechanisms can be implemented.
- •
Scalability, both in bandwidth, network equipment and personnel.
- •
Special events, by incrementing capacity and peak loads for special situations by distributing content as it is needed [13].
CDNs improve performance and availability of web and some media content by pushing the content towards the network edges and providing replication and replica location services. Intelligent replica placement improves response time by serving content from a topological location near the client (in terms of network hops), avoiding the congested backbone networks and network access [14]. Replica location services direct requests for objects to nearby replicas by means of redirections through DNS, based on extensive measurements and monitoring of network performance [15]. The overall performance of a CDN is largely determined by its ability to direct client requests to the most appropriate server [10], [16], [17]. Content providers, such as websites or streaming video sources, contract with commercial CDNs to host and distribute content [18]. They are attractive for content providers because in some cases the responsibility is offloaded to the CDN infrastructure. Most CDNs have servers in ISP points of presence, so clients can access topologically nearby clients with very low latencies. They are capable of sustaining large workloads and flash-crowds due to a large number of servers (Akamai), or few but powerful servers (Digital Island) [11]. The main features of a CDN are:
- •
Decentralizes content storage by moving content closer to clients.
- •
Preserves WAN bandwidth by delivering content locally, and maximizes user performance.
- •
Content management tools help optimize network performance and prioritize mission critical data [19].
CDNs are perfectly integrated in web architecture and the minimum unit managed by them is an object, which are named by URLs. Unlike the web, content providers do not need to manage web servers, since client requests are redirected to replicas hosted by the CDN [12]. CDNs typically host static content (images, advertisements, media clips, etc.) although dynamic content could contain embedded objects served by the CDN [20].
The rest of the paper is structured as follows. Section 2 introduces the motivation and previous work. In Section 3 we present our CDN model starting from previous work. In Section 4 this model is further described and interpreted in an illustrative way for a better understanding. Section 5 presents and explains simulation results obtained from the model. The paper finishes with the conclusions and future work.
Section snippets
Motivation and previous work
CDNs are overlay networks on top of the Internet that deliver content to users from the network edges, thus reducing response time compared to obtaining content directly from the origin server, as depicted in Fig. 1.
If client 1 downloads content from a certain site, it traditionally contacts a centralized server, located in the origin site. The communication may traverse several ISPs and WANs, thus being unable to predict content latency and jitter. If the desired content requires some temporal
The CDN model
The CDN model is composed of three main elements, as shown in Fig. 2.
- –
One origin server, placed at a central location
- –
P surrogates, placed somewhere between clients and the origin server, and
- –
M client clusters, dispersed throughout the world. A client cluster is a way of joining multiple single users located within a certain zone. We will analyze this scenario, obtaining therefore results for the whole group and not for single users. However, our general approach enables the easy extraction of a
Understanding the model
Before translating our expression into simulation scenarios, it could be of interest to better understand it by using some graphical and mathematical techniques.
As there are M client clusters and P surrogates to serve them, one may ask whether P=M, just to associate one optimum surrogate close to each client cluster. However, it is not necessarily true as at the design phase it is extremely difficult to define client clusters. Even so, one may assign various surrogates for the same client
Simulating the model
As can be appreciated, the expression for the main response time corresponds to an n-dimensional function represented with a non-fixed value of n. For example, the probability of a client contacting a surrogate, represented by pij, is really a set of variables whose size depends on the number of clients (M) and surrogates (P). Similar behaviour occurs with latencies (τpi,j,τ0i), capacities (μpj,μs) and mean traffic rates (λi). So we will simulate a CDN for various values of M clients and P
Conclusions and future work
CDNs deal with a communication process where network latency and server capacity are decisive parameters in the response time perceived by the user, as well as the system's client redirection ability to match each client with a suitable nearby surrogate. Performance studies have been made by the research community both in empirical and analytical approaches. This article focuses on the latter approach, starting from previous work where a simple model of a CDN is presented. This paper tries to
Benjamin Molina received his MSc degree in telecommunication engineering from the Universidad Politecnica de Valencia (UPV) in 2001. He made his awarded final project about voice technologies in Tissat, an ICT company in Valencia, where he worked for a year developing PDA web interfaces and CTI services on top of the Java platform. Later he became a member of the Distributed Real-Time Systems research group of the Departamento de Comunicaciones, at the UPV. Benjamin Molina is currently involved
References (27)
- et al.
The trickle-down effect: web caching and server request distribution
Computer Communications
(2002) - et al.
Computer Communications
(2002) - et al.
Selection algorithms for replicated web servers
(1998) Content Distribution Networks, an engineering approach
(2002)- et al.
Analysis of the evolution of peer-to-peer systems
(2002) - et al.
Web caching and content distribution: a view from the interior
(2000) - et al.
An analysis on Internet content delivery systems
(2002) - et al.
Known CDN Request-Routing Mechanisms-draft-cain-cdnp-known-request-routing-01.txt, February
(2001) - et al.
Request redirection algorithms for distributed web systems
IEEE Transaction on Parallel and Distributed Systems
(2003) - Akamai,...
The measured performance of content distribution networks
Globally distributed content delivery
IEEE Internet Computing
Cited by (0)
Benjamin Molina received his MSc degree in telecommunication engineering from the Universidad Politecnica de Valencia (UPV) in 2001. He made his awarded final project about voice technologies in Tissat, an ICT company in Valencia, where he worked for a year developing PDA web interfaces and CTI services on top of the Java platform. Later he became a member of the Distributed Real-Time Systems research group of the Departamento de Comunicaciones, at the UPV. Benjamin Molina is currently involved in research projects related to network simulation environments covering content distribution and scalability issues that may affect real implemented networks. His main interest is focussed on multimedia distribution across Internet and the different related technologies, such as: multicast communications, web caching and real-time systems.
Carlos E. Palau received his MSc and PhD (Dr Ing.) degrees, both in telecommunication engineering, from the Universidad Politecnica de Valencia (UPV) in 1993 and 1997, respectively. He is Associate Professor in the Escuela Tecnica Superior de Ingenieros de Telecomunicación at the UPV, and works in the Distributed Real-Time Systems research group of the Departamento de Comunicaciones. He is currently involved in research and development projects for the application of multimedia and real-time technologies to industry, medicine, education, and communications. Dr. Palau is a member of IEEE and IASTED, and is involved in several IPCs of national and international conferences. He has chaired IASTED Communications Systems and Networks 2002 and 2003.
Manuel Esteve received both his MSc in computer engineering and his PhD in telecommunication engineering (Dr Ing) from the Universidad Politécnica de Valencia in 1989 and 1994, respectively. He is Professor in the Escuela Técnica Superior de Ingenieros de Telecomunicacion at the Universidad Politecnica de Valencia (UPV), and he leads the Distributed Real-Time Systems research group of the Departamento de Comunicaciones. He is currently involved in research and development projects for the application of multimedia and real-time technologies to industry, medicine, education, and communications. He is the responsible for the Virtual University at the UPV And Has Co-Chaired EUROMEDIA'01.