3.4 Tests and Evaluation
3.4.4 Conclusion
In this Chapter we set out to identify whether the Zipf model or Zipf-Mandelbrot model shares a greater likeness to VoD observed request distributions. Two empirical VoD request distributions were supplied by BT to aid in identifying the most appropriate model. The methods used for measuring were Pearson Chi-Squared and KullBack- Leibler which each provided a best representative distribution of each model tested. Each representative was compared using the “goodness-of-fit” measurement produced by each tests. Each method of testing used deduced that Zipf-Mandelbrot is indeed a better representative than Zipf in relations to the VoD and TV catch-up video request distributions.
Additionally to the KL and PCS methods of testing and comparing, an ICN simula- tion environment with cache enabled nodes was used to measure the network behaviour of the empirical data in relation to the KL and PCS deduced best representative dis- tributions from each model. The outcome from these tests highlights the measured characteristics (average latency, average path stretch and average cache-hit ratio) from the Zipf-Mandelbrot distributions share a greater likeness to the empirical VoD request distributions’ measured characteristics when compared to the Zipf distributions. Con- cluding, Zipf-Mandelbrot from more closely modelled, the empirical data in terms of expected network behaviour in an ICN cache enabled network.
4
VoD Request Simulation Environment
A pseudo-realistic video request generator would provide valuable insight into video delivery systems. The target of identifying the spectrum of parameters required to sim- ulate such an environment applying pseudo-realistic parameters that would accurately estimate the request behaviour for a specific demographic is challenging. The parame- ters can be identified in video delivery systems currently already deployed or through surveys of a specific area / groups identified as the target audience. The parameters would include such things as: the quantity of videos at any time; lifetime of videos on your system; decay of video popularity of your system; distribution of video popularity; daily / weekly distribution of request relative to time of day and the number of total active users of your system. These are just some of the variables that go into generating realistic user requests. This Chapter sets out to design and implement such a VoD request generator accounting for all the above parameters.
4.1
Introduction & Motivation
The ability to simulate video requests with reasonable accuracy would provide an asset to the Video on Demand market that has become the majority of internet traffic in present days [1, 6]. A video request simulator would equip video content providers with a tool to predict, and thus resolve, networking and application platform issues before they arise when the cause of the issues stem from the design of the system in question. In Chapter 5 a cache eviction algorithm is introduced for which a request generator, such as the one in this Chapter, would provide valuable insight into the behaviour of such an algorithm.
Users of Video on Demand systems appear not to exemplify a drastic range of behaviours, though still complicated as many environmental variables are at play at any one moment. The increase of popularity per item when introduced, the decrease of popularity of items once their maximum has been reached, the cumulative request total per item, the request quantities throughout any given time are all factors that can be broken down to be included in a single simulator which is what is attempted in this chapter.
One limitation of the pursuit to produce a simulator that encompasses all the vari- ables mentioned is the lack of quantifiable research data that can be used to produce such a simulator. A number of research papers have attempted to find and quantify the parameters required, however, they appear to forget to consider the problem in its entirety, which limits the possibility of simulating VoD data accurately. An example of this is the decay of video objects which has been researched [16,36,37], however, without also considering growth of popularity video objects, this research is incomplete.
Popularity distributions of VoD systems are also important to consider and are found in many publications [9,10,14,17,18,22,33,38,70,71]. They provide a brief glance (typically ranging from a week to a number of weeks) into observations of the popularity of all items in a VoD system (often found to be approximately Zipf, however debated to be Zipf-Mandelbrot in Chapter 3).The primary limitations of these observations are
that new items are introduced in this period and items entered prior to the start of the observation are only observed in their decayed state. The popularity distribution not measured is the total request distribution for all items at a single moment in time from their conception to removal (“birth” and “death”). Such a probability distribution would provide one with the request distribution of all items at a single moment in time. The items observed may be in the system for various lengths of time, however, it would give a more complete picture of the popularity distribution one can expect.