SVC Traffic and Quality Modelling

Introduction

Due to the usage of layers and hierarchical B-frames in SVC, scalable streams differ from ones encoded with H.264/AVC in structure of groups of pictures (GoPs). Therefore, the properties of SVC streams differ significantly from streams encoded with H.264/AVC. SVC streams feature high bitrate variability and non-negligible autocorrelation of GoP lengths.

We propose an innovative approach to joint modelling of traffic patterns and quality of SVC streams. Comprising two independent components — i.e., the traffic generator and the quality estimator — the model can significantly speed up the development of SVC applications in various use-cases by relying on approximated properties of streams instead of real traces. The model is verified based on a set of video streams and the performance of its components is assessed.

Components

The proposed model consists of two autonomous and independent components: the traffic generator and the quality estimator. These may be run separately or together. The process of estimation of their parameters is presented below.

Components

Verification

The model was implemented and verified in MATLAB. To gather the training data that resemble the content being streamed by centralized and P2P-based Internet television services, 19 high quality (HD 720p) movies of different lengths (feature films, serial episodes, short films) and genres (thriller, documentary, animation, sci-fi etc.) were downloaded from www.vodo.net:

Depending on the case, 400 to 1,200 sequences (covering 10% to 30% of total content length) were randomly selected to train the components. They were encoded by the SVC JSVM reference software codec into three layers: base layer L0 (with higher quantization parameter QP) and enhancement layer (with lower QP) split into two MGS layers L1 and L2. Each sequence contained 15 GoPs with G16B15 pattern (one I~frame and 15 B frames per GoP). Both traffic generator and quality estimator were trained using the same sequences.

The cumulative distribution functions (CDF) of the GoP lengths used for training and generated ones are compared below.

svc-traffic-and-qoe-fig3

Below, the goodness of fit of autocorrelation function of GoP lengths is presented.

Verification - ACF

These two figures prove that the traffic generator component is able to approximate the probability distribution and the autocorrelation perfectly at the same time.

The figure below depicts the goodness of fit of the quality estimator’s parameters using the most intuitive description of the stream — i.e., the mean bitrate over some period of time (in this case, over the time the video is buffered for VQMcomputation).

Verification - VQM

As can be seen from the scatter and contour plots, the model can be successfully fitted to the training set and the generated correlation resembles the training one.

Applicability and Performance

As presented below, three distinct use-cases have been identified to prove the applicability of the solution:

  • ISPs may use the traffic generator to test their traffic management solutions without copying real streams or re-playing traces;
  • content providers may use the quality estimator to assess the quality of the streams they disseminate;
  • research centers may use both components to simulate the behavior of their solutions at all development and deployment stages.

svc-traffic-and-qoe-fig6

To assess the performance in a non-MATLAB environment, the components have been implemented in Java and the values obtained in MATLAB during the verification were used. The implementation’s performance was assessed on two computers: a server and a PC. The server was equipped with two quad-core Intel Xeon W5590 CPUs running at 3.33 GHz and 48 GiB of RAM. It was running CentOS 5.9 and Oracle JRE 6. The PC was equipped with one dual-core Intel Core i3 540 CPU running at 3.07 GHz and 4 GiB of RAM. It was running Ubuntu 12.04 and Oracle JRE 7. In each use-case and computer combination, the implementation was run independently 30 times, each time generating 4,050,000 samples (30 days of 25 fps stream). The table below presents two measures of performance: the mean number of GoPs per second generated and the mean speedup, i.e., the ratio of the simulated stream length to the simulator’s working time.

GoPs per second speedup
server  PC server PC
traffic generator 218 k 165 k 140 k 105 k
quality estimator 10.1 k 8.60 k 6.47 k 5.50 k
generator + estimator 9.79 k 8.33 k 6.27 k 5.33 k

As can be seen, the speedup of all components is significant and allows processing of huge amounts of data in a short time.

Publications

The model is presented in:

  • P. Wydrych, K. Rusek, and P. Chołda, “Efficient modelling of traffic and quality of scalable video coding (SVC) encoded streams,” IEEE Communications Letters, vol. 17, no. 12, pp. 2372–2375, Dec. 2013. ©IEEE, available on IEEE Xplore.

    In this letter, a new model for generating SVC-like traffic and estimating its quality is introduced. Comprising two independent components — i.e., the traffic generator and the quality estimator — the model can significantly speed up the development of SVC applications in various use-cases by relying on approximated properties of streams instead of real traces. The model is verified based on a set of video streams and the performance of its components is assessed.

Source Code

When using the source code, please cite the abovementioned publication.