Elsevier

Computer Communications

Volume 29, Issue 12, 4 August 2006, Pages 2322-2347
Computer Communications

A performance comparison of self-organising application layer multicast overlay construction techniques

https://doi.org/10.1016/j.comcom.2006.02.020Get rights and content

Abstract

Application layer multicast (ALM) uses overlays built on top of existing network infrastructure for rapid deployment of multicast applications. Key to the efficiency of this technique is the structure of the overlay tree used. This work reviews and compares various self-organising techniques that strive to build low cost, and low delay trees using extensive simulations. Protocols investigated include HMTP, HostCast, switch-trees, DCMALTP, NICE, TBCP and Narada which encompass a wide spectrum of overlay construction, optimisation and maintenance techniques. The protocols are evaluated based on their ability to achieve their objectives, overlay path penalties, protocol convergence and overhead. We also conduct detailed analysis of two main components in building an overlay: initial construction and the overhead of periodical improvement. Based on the observed results, we identify strengths and weaknesses of various approaches, and provide suggestions for future work on ALM overlay optimisation.

Introduction

Multicast provides a cost-effective data delivery mechanism for group communication. The lack of globally available IP multicast infrastructure has led to an increasing interest in application layer multicast (ALM) [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17]. ALM provides multicast primitives, i.e., packet replication, at the end systems, and uses existing unicast forwarding service for data distribution. These forwarding paths form a logical topology, or overlay on top of the network layer infrastructure.

The simplest overlay model for group communication is a set of independent direct unicast connections from the data source to all other participants. This however creates a huge amount of redundant traffic in the network and overloads the source node. These problems are aggravated with a growing number of participants. ALM overlays alleviate this by organising members into a delivery tree such that each member only needs to forward data to a small set of other members. Hence, the key to the efficiency of an ALM solution lies in the tree structure used. This paper focuses on various approaches that aim to build “good” ALM overlays.

A major problem in building optimised overlays is often the end-system’s lack of knowledge of the underlying topology and its network metrics, such as delay and bandwidth. To infer such information, an ALM protocol needs to use end-to-end measurement techniques. However, it is impractical to gather the metrics for all node-pairs in an n-node overlay as it requires O (n2) measurements. In other words, an ALM solution has to work with limited topology information. In addition, limited access bandwidth at end systems restricts the branching degree (or fan-out) of a delivery tree. Unfortunately, many degree-constrained spanning trees optimisation problems (e.g., minimum cost or delay) have been proven to be NP-hard [18], [19], [20].

We focus on existing distributed ALM protocols that utilise limited network measurements to construct and optimise the overlays. Two widely studied performance goals are considered here: minimising the delivery tree’s cost and minimising the end-to-end delivery latency. These two objectives are important for applications such as bulk data distribution and real-time media streaming.

Protocols studied in this paper are: HMTP [1], HostCast [10], switch-trees [21], DCMALTP [8], NICE [13], TBCP [12] and Narada [7]. They represent a wide variety of choices in self-organising overlay construction (tree-first and mesh-first), data delivery (shared spanning tree and source-specific trees), maintenance (tree-based, mesh-based and hierarchical cluster-based) and optimisation (fully distributed and localised central computation model). Observations from this diversity of combinations provide insight for future ALM overlay research.

We use extensive simulations to study various aspects of the above protocols. To the best of our knowledge, there has been no attempt to compare all the above-mentioned protocols in an identical environment. Our results show the important performance differences and similarities between the protocols in terms of optimisation metrics, the topologies formed and their protocol convergence and overhead. A survey of this size cannot cover every aspect of every protocol and we justify our choice of comparisons throughout the paper.

The rest of the paper is organised as follows. Section 2 provides a discussion of existing tree cost- and delay-optimised overlay construction proposals. Section 3 presents our evaluation settings, results and analysis. Section 4 describes related work. Finally, Section 5 concludes the paper along with some future research directions.

Section snippets

Cost- and delay-optimised overlay construction techniques

We begin the discussion by describing problems in tree cost and delay optimisations in the next subsection. Subsequent section presents the general working of an ALM overlay construction process, which is followed by a review of protocols considered in this paper.

Simulation setting

We developed a java-based packet-level, discrete event simulator for the evaluations. The simulator assumes shortest path routing between any two nodes in the network. The network is represented by a set of routers and links. The simulator does not model either queuing delay or packet losses to facilitate large-scale evaluations.

We use topologies ranging from 600 to 3040 nodes from three distinct topology models:

  • Waxman model. This model randomly assigns nodes to locations on a plane, and

Other ALM protocols

This work focuses on tree cost- and delay-optimised overlay construction techniques. Another widely studied performance objective is bandwidth. Overcast [11] and SpreadIt [16] are designed for one-to-many bandwidth-intensive applications. Overcast adopts the switch-1hop technique (see Section 2.3.1.2), while SpreadIt uses a scaled-down version of TBCP to achieve its objectives. Hence, we believe that some of our findings are applicable to these protocols. Yoid [17], one of the earliest ALM

Conclusion and future work

In this paper, we investigate the efficiency of several self-organising techniques to build low cost and low delay application layer multicast (ALM) trees. The studied techniques encompass representatives from two main overlay construction techniques, i.e., tree-first and mesh-first. Tree-first protocols considered include HMTP, TBCP, NICE, and variants of switch-trees, HostCast and DCMLTP. We use Narada as the mesh-first representative. We evaluate various aspects of these protocols under an

Su Wei, Tan is currently a lecturer at the Faculty of Engineering, Multimedia University, Malaysia. His research interest includes application layer multicast, peer-to-peer and overlay networks, and quality of service in wireless networks. Su Wei holds B.Eng (Electrical), M.Eng.Sc and PhD degrees from University of Malaya (Malaysia), Multimedia University (Malaysia) and University of Kent (UK) respectively.

References (46)

  • B. Zhang, S. Jamin, L. Zhang, Host Multicast: a framework for delivering multicast to end users, in: IEEE Infocom, New...
  • D. Pendarakis, S.Y. Shi, D. Verma, M. Waldvogel, ALMI: an application layer multicast, in: 3rd USENIX Symposium on...
  • J. Liebeherr, M. Nahas, Application-layer multicast with delaunay triangulations, in: IEEE Global Internet Symposium...
  • S. Ratnasamy, M. Handley, R. Karp, S. Shenker, Application-level multicast using content-addressable networks, in:...
  • D.A. Helder, S. Jamin, Banana Tree Protocol: An End-host Multicast Protocol, University of Michigan, Technical Report...
  • S. Zhuang, B. Zhao, A. Joseph, R. Ratz, J. Kubiatowicz, Bayeux: an architecture for scalable and fault-tolerant...
  • Y.H. Chu, S.G. Rao, H. Zhang, A case for end system multicast, in: ACM SIGMETRICS, Santa Clara, CA, 2000, pp....
  • S. Banerjee, C. Kommareddy, K. Kar, B. Bhattacharjee, S. Khuller, Construction of an efficient overlay multicast...
  • V. Roca and A. El-Sayed, A host-based multicast (HBM) solution for group communication, in: IEEE International...
  • Z. Li, P. Mohapatra, HostCast: a new overlay multicast protocol, in: IEEE International Conference on Communications...
  • J. Jannotti, D. DGifford, K. Johnson, M. Kaashoek, J. O’Toole, Overcast: reliable multicasting with an overlay network,...
  • L. Mathy, R. Canonico, D. Hutchison, An overlay tree building control protocol, in: Networked Group Communication,...
  • S. Banerjee, B. Bhattacharjee, C. Kommareddy, Scalable application layer multicast, in: ACM SIGCOMM, Pittsburgh, PA,...
  • S. Jain, R. Mahajan, D. Wetherall, G. Borriello, Scalable Self-organizing Overlays, University of Washington, Seattle,...
  • M. Casto et al.

    Scribe: a large-scale and decentralized application-level multicast infrastructure

    IEEE JSAC

    (2002)
  • H. Deshpande, M. Bawa, H. Garcia-Molina, Streaming Live Media over a Peer-to-Peer Network, Stanford University...
  • P. Francis, Yoid: Extending the Internet Multicast Architecture, Unpublished report, Available from:...
  • J. Konemann, Approximation Algorithms for Minimum-cost Low-degree Subgraphs, PhD Thesis (Computer Science). Carnegie...
  • S.Y. Shi, J.S. Turner, M. Waldvogel, Dimensioning server access bandwidth and multicast routing in overlay networks,...
  • N.M. Malouch, Z. Liu, D. Rubenstein, S. Sahu, A graph theoretical approach to bounding delay in proxy-assisted,...
  • D.A. Helder, S. Jamin, End-host multicast communication using switch-trees protocols, in: Workshop on Global and Peer...
  • T.H. Cormen et al.

    Introduction to Algorithms

    (1991)
  • A. El-Sayed et al.

    A survey of proposals for an alternative group communication service

    IEEE Network

    (2003)
  • Cited by (17)

    • ALMware: A middleware for application layer multicast protocols

      2011, Computer Communications
      Citation Excerpt :

      Moreover, the common environment enables us to build and run several ALM implementations and compare them. Performance comparison of ALM protocols by simulation has been conducted [24,25] and dedicated simulators have been provided [26]. However, to our best knowledge, no paper has reported the comparison of ALM protocols in real environments under common settings, and no tools and methodologies for such purpose have been presented so far.

    • An application layer multicast approach based on topology-aware clustering

      2009, Computer Communications
      Citation Excerpt :

      In the implicit approach, the mesh and tree are simultaneously defined, such as [3,9,8]. Many application layer multicast protocols (e.g. [3,2,10]) employ distributed depth first searching approach to position a host (see [16]). In DFS, a newcomer searches down the multicast tree by exploring the branches of some existed nodes.

    • A Centralized Optimization Solution for Application Layer Multicast Tree

      2017, IEEE Transactions on Network and Service Management
    • Multi-domain clustering topology-aware application layer multicast

      2015, Huazhong Keji Daxue Xuebao (Ziran Kexue Ban)/Journal of Huazhong University of Science and Technology (Natural Science Edition)
    View all citing articles on Scopus

    Su Wei, Tan is currently a lecturer at the Faculty of Engineering, Multimedia University, Malaysia. His research interest includes application layer multicast, peer-to-peer and overlay networks, and quality of service in wireless networks. Su Wei holds B.Eng (Electrical), M.Eng.Sc and PhD degrees from University of Malaya (Malaysia), Multimedia University (Malaysia) and University of Kent (UK) respectively.

    Gill Waters is an Honorary Senior Lecturer in Computer Science at the University of Kent, UK. Her research at Kent concerns distributed multicast applications and the required network architecture, protocol and system support for these applications. This has included multicast routing with bounded delays and incorporating performance and validation considerations in the design of distributed systems, projects funded by the UK’s Engineering and Physical Sciences Research Council and by British Telecom. Since retiring from her permanent post in 2005, she is concentrating on protocol design for network structures, using distributed algorithms inspired by a variety of techniques for clustering. Gill was a lecturer in the Department of Electronic Systems Engineering at the University of Essex (1984-1994) and previously worked at the Rutherford Appleton Laboratories on a wide range of innovative computer communication projects. Gill has a BSc in Mathematics (University of Bristol) and a PhD in Electronics (University of Essex).

    John Crawford is a Lecturer in Computer Science at the University of Kent. He has an industrial background in the development of communications systems, where he has held roles ranging from software engineer to consultant. His research interest concerns multicast routing algorithms, protocols and quality of service issues. John holds MSc and PhD degrees from the University of Kent.

    View full text