INTRODUCTION

The increasing number of connected displays is driving the current explosion of Internet video traffic. In this context, streaming over HTTP has become the dominant approach for delivering multimedia content over the Internet.

Several versions of HTTP adaptive streaming have been proposed by various stakeholders and pushed for the provision of “over the top” audiovisual delivery in the Internet.

In addition, standardization efforts have been made. Since 2012, the MPEG Dynamic Adaptive Streaming over HTTP [1] (MPEG- DASH) standard is available. Such adaptive protocols will be referred to as “HTTP Adaptive streaming” or HAS throughout this paper.

While these techniques have been designed to cope with varying delivery conditions, under some specific challenging circumstances, HAS client implementations suffer from severe performance issues, namely instability, unfairness and bandwidth underutilization.

The first part of the work is dedicated to HAS competition with a greedy TCP flow. We propose a central dynamic traffic shaping method to mitigate the convergence issue and to optimize HAS client convergence.

The second part addresses HAS competition among several clients through implementing a distributed algorithm. We define information to be exchanged between clients and a common set of rules that enables each client to determine the best representation to be requested.

The benefits of this solution will be evaluated thanks to an implementation based on MPEG-DASH standard [1].

Finally, the paper presents the standardization effort made to include this technique into MPEG-DASH.

RELATED WORK

The main idea of HAS techniques is to ensure a continuous playback by choosing representations whose bit-rate fit the average available bandwidth.However this available bandwidth is generally only estimated by the client from the received rate of the content.

When several clients receive data at the same time, TCP allows to provide a fair share of the bandwidth to the clients. But since HAS content is segmented, each client has periods of ‘full speed’ download interleaved with periods of no activity.

When several clients operate on the same local network, the way these (variable) periods overlap influences the perceived TCP throughput. The clients may see instable measured bandwidth which leads to potential user experience degradation such as unstable choice of representation or unbalanced share of bandwidth between competing clients.

This issue, known as the ‘downward quality spiral phenomenon’, has been characterized in [2] and explained in [3] and [4]. To mitigate it, two strategies may be used: improve the HAS implementation or manage the content delivery from the network infrastructure.

Regarding the HAS implementation first, [4] and [8] proposed some improvements at the client side. The main idea consists in reducing the client conservatism. The client becoming greedier has a better chance of grabbing bandwidth. Nevertheless greedy clients overreact to the bandwidth variations [10].

A second idea, investigated in [8], proposes to modify the segment request scheduling to randomize the ON-OFF periods and then break the downward spiral.

Implementation improvements have also been proposed on the server side. S. Akhshabi and al. 9 try to tackle the instability issue modifying the HAS server to detect oscillating clients and to shape the rate of their requested segments to reduce as much as possible the OFF period duration causing the instability.

Nevertheless to work correctly, all the segments should be handled by the same server which is quite restrictive with regard to CDN infrastructures.

DOWNLOAD THE FULL TECH PAPER BELOW