Scalability evaluation of the Yima streaming media architecture

Authors

  • Roger Zimmermann,

    Corresponding author
    1. Integrated Media Systems Center and Department of Computer Science, University of Southern California, Los Angeles, CA 90089–2561, U.S.A.
    • Integrated Media Systems Center and Department of Computer Science, University of Southern California, Los Angeles, CA 90089–2561, U.S.A.
    Search for more papers by this author
  • Cyrus Shahabi,

    1. Integrated Media Systems Center and Department of Computer Science, University of Southern California, Los Angeles, CA 90089–2561, U.S.A.
    Search for more papers by this author
  • Kun Fu,

    1. Integrated Media Systems Center and Department of Computer Science, University of Southern California, Los Angeles, CA 90089–2561, U.S.A.
    Search for more papers by this author
  • Shu-Yuen Didi Yao

    1. Integrated Media Systems Center and Department of Computer Science, University of Southern California, Los Angeles, CA 90089–2561, U.S.A.
    Search for more papers by this author

Abstract

Over the last decade research has been pursued on all aspects of streaming media. While many theoretical results have been reported in the literature, few performance results of large-scale systems have been published. In this report we specifically explore the scalability aspects of our Yima streaming media architecture in an end-to-end test environment. With Yima, it was our goal to design and implement an architecture that would scale in performance from small to large systems. Some of the design features include (1) a multi-node cluster architecture based on commodity hardware and custom software, (2) media type independence (support ranges from 500 Kb sequation image MPEG-4 to 45 Mb sequation image HDTV, at both variable and constant bitrates), (3) fine-grained online scale up/down capabilities, and (4) a client-controlled rate smoothing protocol. We briefly discuss the design and implementation of these capabilities of Yima and then thoroughly evaluate its scalability through several sets of experiments. Our results show that Yima scales linearly (within the range of our test parameters) as a function of the cluster size and also as a function of available resources such as network bandwidth and CPU performance. Copyright © 2004 John Wiley & Sons, Ltd.

Ancillary