Resource discovery and request-redirection for dynamic load sharing in multi-provider peering content delivery networks

https://doi.org/10.1016/j.jnca.2009.03.003Get rights and content

Abstract

A constellation of Content Delivery Networks (CDNs), termed as peering CDNs, endeavors to guarantee adequate delivery performance when the incoming request load is overwhelming for a single provider alone. Each user is served by an optimal Web server in terms of network cost, even under heavy load conditions. Before it could be comprehended, appropriate resource discovery and request-redirection mechanisms, coupled with an optimal server selection strategy, should be in place to perform the distribution of highly skewed loads. In this paper, we devise an effective load distribution strategy by adopting distributed resource discovery and dynamic request-redirection mechanisms, taking traffic load and network proximity into account. The load distribution strategy reacts to overload conditions, at a time instance, in any primary CDN server(s) and instantly distributes loads to the target servers, minimizing network cost and observing practical constraints. In this context, we exercise an asynchronous resource discovery protocol, reminiscent of the public/subscribe notion, and formulate the resulting redirection scheme. Extensive simulation analyses demonstrate the novelty of our approach. In particular, we show that our approach is effective to handle high load skews by preserving locality, and thus achieve service “responsiveness”. We also perform a sensitivity analysis to reveal that our redirection scheme outperforms other alternatives to handle peak loads.

Introduction

Content Delivery Networks (CDNs) (Buyya et al., 2008; Pallis and Vakali, 2006) evolved as a solution of Internet service degradations such as congestions and bottlenecks due to the large end-user demands posed on Web access services. To operate effectively, often a CDN is required to break down system silos to increase utilization rates, either through over-provisioning its capacity or harnessing external resources on demand. The requirements for providing high-quality service through global coverage can be fulfilled through a constellation of CDNs, termed as ‘peering CDNs’ (Pathan et al., 2007). Such collaboration, leveraging existing infrastructures, is not only important from a reachability perspective but also from quality and performance perspective. Peering between CDNs can be observed for a short- or long-term duration to handle workload variations.

The success of peering and the effectiveness of operations for content delivery in a peering CDNs system depend on its ability to perform resource discovery, server selection and dynamic request-redirection under degenerated load conditions (e.g. flash crowds). The resource discovery process specifies how external resources offered by disparate CDNs are discovered. An effective server selection strategy determines the optimally underloaded edge server(s) that is best suited to serve user requests. The server selection phase typically chooses the “nearest” optimal server to the requesting user. A dynamic request-redirection mechanism assists in directing user requests to the target edge server(s), so as to alleviate imbalanced load situations. These phases may be interleaved to collectively perform load distribution by reacting to overload conditions in a multi-provider peering CDNs system and thus endeavor to achieve scalability.

Many previous research (Cardellini et al., 1999a, Cardellini et al., 1999b; Colajanni et al., 1998; Conti et al., 2001; Harchol-Balter et al., 1999; Lamnitchi and Foster, 2003; Schwartz et al., 1992; Shnayder, 2003; Zhu and Hu, 2005) have focused on devising resource discovery and redirection algorithms for distributed Web servers, overlay networks, Internet, large-scale Grids, and Peer-to-Peer (P2P)-based systems. However, they cannot be directly applied for load balancing in peering CDNs, due to the necessity for handling dynamic circumstances, thus requiring up-to-date information about widely distributed resources. In addition, providers should learn about available resources quickly, without using an inordinate amount of communication, and the resource discovery and redirection algorithms may be used repeatedly to obtain updated resource status information. There are also other challenges, which include virtualization of multiple providers and offloading requests from the overloaded provider to its underloaded peers, based on cost, performance and load. In such a cooperative multi-provider environment, requests are directed to sets of servers deployed across multiple CDNs as opposed to individual servers belonging to a single entity. Therefore, resource discovery and request-redirections must occur over distributed sets of servers spanning multiple CDNs, without having complete state information.

In this paper, we present distributed resource discovery and dynamic request-redirection algorithms for an effective load distribution strategy. Our aims are: (i) to perform dynamic load distribution under traffic surges by redirecting excess requests to optimally underloaded Web server(s), thus binding users to optimal replicas (timeliness); (ii) to exhibit acceptable throughput under overload conditions (e.g. during flash crowds); (iii) to scale to distributed inter-CDN resources scattered across the globe (dynamic lookup); and (iv) to maintain administrative control over local resources and their states (resource encapsulation).

Specifically, the communication protocol to aid resource discovery conservatively implements the public/subscribe paradigm. The use of the public/subscribe notion endeavors to perceive—scalability and full decoupling from other system operations; a possibly “offline” approach due to the asynchronous nature of resource discovery; and indirect addressing for load balancing. At the heart of the load distribution strategy lies the redirection scheme that takes traffic load and network proximity into account. In our approach, load indices of distributed inter-CDN servers are obtained through an asynchronous feedback mechanism and network proximity is measured using a pinger logic with low messaging overhead. A simulation model, capturing key system components, is developed to evaluate the performance of our approach. Experiment results reveal that an acceptable level of throughput can be achieved, even under heavy load, and the proposed redirection scheme outperforms other alternatives. The main contributions of this paper are:

  • An asynchronous resource discovery algorithm without any central coordinating authority to identify resources from disparate CDNs.

  • A load and proximity-aware request-redirection algorithm that reacts to overloaded server conditions in multi-provider peering CDNs by steering excess user requests to optimally underloaded servers.

  • A comparison-based simulation analysis to evaluate the performance and perceived benefits of our approach, and a sensitivity analysis of the proposed redirection scheme using critical system parameters.

The rest of the paper is structured as follows. In Section 2, a brief description of the peering CDNs is provided. It is followed by the proposed resource discovery and request-redirection algorithms. Simulation methodology is described in Section 4 and results are presented in Section 5. A comparative analysis of our approach to the existing work is followed next. Finally, Section 7 concludes the paper.

Section snippets

Peering CDNs: overview

A peering arrangement is a conceptual layer, i.e. overlay, over the physical CDN networks, which play the main role by establishing agreements to share peers’ resources, and by cooperating to create a rich computational environment for effective content delivery. The initiator of a peering negotiation is called a primary CDN; while other CDNs who agree to provide their resources are called peering CDNs or peers. Resources belonging to a peering arrangement of CDNs are scattered over the globe,

Algorithms

In this section, we describe the workings of the resource discovery protocol in peering CDNs. We also provide the description of the load and proximity-aware request-redirection strategy, coupled with optimal server selection, in order to perform dynamic load distribution in peering CDNs.

Methodology

Measurement-based performance studies could be advantageous and suitable when a real system testbed or prototype is available. However, they may not reproduce the problems and scenarios for which the solutions are designed, since in real testbeds several important parameters, such as server and network load conditions, cannot be controlled. In addition, it is extremely difficult to have a significant amount of geographically dispersed end-users simultaneously to generate traffic causing a flash

Experimental evaluation

In this section, simulation results are presented to evaluate the performance and to provide critical assessment of our approach. We run our experiments for the reference simulation model of Section 4, with one provider as primary (CDN 1) and others as peers. Results are obtained from ten simulation runs, where each run is for duration of 10,000 s (approx. 3 h) of the peering CDNs system activities. While our simulations are designed to converge to the “true solution” of the model, running the

Related work

Resource discovery is a popular topic in large-scale distributed systems; whereas, request-redirection is an indispensable enabling cornerstone for CDNs. Many research efforts have focused on these two topics separately in different domains, such as Grid computing, P2P-based systems, overlay networks, multi-agent systems, and ad-hoc networks. Analyses of previous research efforts, in relation to content internetworking, suggest that there has been only modest progress on dealing with resource

Conclusion and future works

In this paper, we present resource discovery and request-redirection algorithms for a dynamic load distribution strategy, which alleviates any load imbalance in a peering CDNs system. Resource discovery in our approach follows a distributed and asynchronous nature, using a communication protocol that conservatively implements the public/subscribe paradigm. In addition, request-redirection occurs over distributed sets of servers, minimizing redirection cost. Specifically, in our approach, when

Acknowledgements

We are indebted to Christian Vecchiola for his selfless help and support during the simulation setup and experiments. We would like to thank Marcos Assunção, SungJin Choi, Mustafizur Rahman, Rajiv Ranjan and Saurabh Garg from the University of Melbourne, Australia for sharing thoughts and for making incisive comments and suggestions on this paper. We are also thankful to anonymous reviewers for their valuable and constructive comments to improve the paper's structure, quality and readability.

References (30)

  • A. Biliris et al.

    CDN brokering

    Computer Communications

    (2002)
  • R. Chand et al.

    Powerful resource discovery for Arigatoni overlay network

    Future Generation Computer Systems

    (2008)
  • H.A. Alzoubi et al.

    Anycast CDNs revisited

  • L. Amini et al.

    Modeling redirection in geographically diverse server sets

  • Barbir A, Cain B, Nair R, Spatscheck O. Known content network (CN) request-routing mechanisms. RFC 3568, July...
  • Broberg J, Tari Z. MetaCDN: Harnessing storage clouds for high performance content delivery. In: Proceedings of the 6th...
  • V. Cardellini et al.

    Dynamic load balancing on web-server systems

    IEEE Internet Computing

    (1999)
  • Cardellini V, Colajanni M, Yu PS. Redirection algorithms for load sharing in distributed Web-server systems. In:...
  • Colajanni M, Yu PS, Cardellini V. Dynamic load balancing in geographically distributed heterogeneous Web servers. In:...
  • M. Conti et al.

    QoS-based architectures for geographically replicated web servers

  • M.E. Crovella et al.

    A self-similarity in world wide web traffic: evidence and possible causes

    IEEE/ACM Transactions on Networking

    (1997)
  • M.E. Crovella et al.

    Heavy-tailed probability distributions in the world wide web

  • Day M, Cain B, Tomlinson G, Rzewski P. A model for content internetworking. IETF RFC 3466, February...
  • Ercetin O, Tassiulas L. Request routing in content distribution networks. Technical report, Sabanci University, Turkey,...
  • Cited by (19)

    • Resource discovery for distributed computing systems: A comprehensive survey

      2018, Journal of Parallel and Distributed Computing
      Citation Excerpt :

      In synchronous methods [142,187,20,74,65], the search algorithm proceeds synchronously, in parallel rounds, where each round is defined as the time required for each node in the network to communicate with one or more other nodes, learning about them and gathering information. Asynchronousmethods [3,186,238] are based on the asynchronous communication model, where each node can send a message in arbitrary size to any of its neighbors in a way that the outgoing message eventually arrives in the destination node after an unbounded finite time. Moreover, unlike the synchronous method, there is no assumption to start the algorithm simultaneously on all the nodes, and the receiving nodes simply follow the First-Input First-Output (FIFO) model to serve the requesters.

    • Fractal: An advanced multidimensional range query lookup protocol on nested rings for distributed systems

      2017, Journal of Network and Computer Applications
      Citation Excerpt :

      One of the major unsolved problems in distributed systems is the support of efficient multi-attribute range queries. Several methods have attempted to solve this problem in their lookup methods (Souri and Navimipour, 2014; Caminero et al., 2013; Carlini et al., 2016; Lee et al., 2013; Pathan and Buyya, 2009; Paganelli and Parlanti, 2012; Cai et al., 2004; Shen and Xu, 2012; Schmidt and Parashar, 2008; Jagadish et al., 2005; Pitoura et al., 2012; Arefin et al., 2009), but ignoring the vital characteristics of methods proposed in large-scale distributed systems such as P2P, grid, and IoT (Atzori et al., 2010) makes these methods impractical in out-laboratory environments. One of the most popular and successful solutions for the service discovery problem is the use of Distributed Hash Tables (DHTs) (Lee et al., 2013; Paganelli and Parlanti, 2012; Cai et al., 2004; Shen and Xu, 2012; Schmidt and Parashar, 2008; Carlini et al., 2011, 2010; Albrecht et al., 2008; Vasilakos et al., 2015; Xu et al., 2010).

    • Space4time: Optimization latency-sensitive content service in cloud

      2014, Journal of Network and Computer Applications
      Citation Excerpt :

      In most of the existing works, requests are simply forwarded to the upper-layer parent servers when the content is not locally available. Or request-redirection occurs over distributed sets of servers, to minimize redirection latency (Pathan and Buyya, 2009). This adaptation may include dynamically shipping the request as well as the requested data set to some vantage locations in the cloud that are close to the users so that the total access latency of the request sequence and migration costs of the request and data are minimum.

    • Optimization on content service with local search in cloud of clouds

      2014, Journal of Network and Computer Applications
    • ANFIS and agent based bandwidth and delay aware anycast routing in mobile ad hoc networks

      2014, Journal of Network and Computer Applications
      Citation Excerpt :

      To provide load balancing among the sinks, it proposes a sink selection mechanism based on fuzzy logic for the frame forwarding which evaluates the traffic density in the direction of each sink by combining two dynamic criteria which are the number of contenders and the buffer occupancy levels in the neighborhood with the static distance criterion. Load distribution strategy by adopting distributed resource discovery and dynamic request-redirection mechanisms by using anycast, considering traffic load and network proximity is discussed in Mukaddim and Rajakumar (2009) for content delivery network servers. A distributed algorithm for sink selection in wireless sensor network is discussed in Trivino et al. (2011).

    View all citing articles on Scopus
    View full text