Practical Evaluation of Multi-Source Coded Downloads

Abstract

In this paper, we introduce two multi-source download protocols for loosely orchestrated networks that have high potential in Information-Centric Networking (ICN). We focus on services with high bandwidth and low delay requirements, such as video streaming. We propose MUlti-source Transmission Protocol (MUTP) for uncoded multi-source data delivery and extend it with network coding capabilities to create Coded MUTP. We investigate their throughput using a custom-designed system that includes browser extensions and proxy servers. The browser extensions intercept YouTube video downloads and forward them through our proxy server, using parallel HTTP requests, Uncoded MUTP or Coded MUTP approach. We present measurement results collected in 2018-2019, over eleven months that include 1,300,000 log records from more than 960 GBs of video download. We show that even when downloading from only two sources, our protocols can match the heavily optimized HTTP. Furthermore, by increasing the number of sources to four or higher, MUTP protocols can outperform HTTP, reaching an up to three-fold goodput (useful throughput) increase. In addition, we show that the proposed solution avoids the straggler problem, therefore adding more sources to a network increases its goodput.