DecVi: Adaptive Video Conferencing on Open Peer-to-Peer Networks

TL;DR

DecVi is a decentralized protocol for constructing efficient multicast trees in open peer-to-peer video conferencing networks, improving reliability and quality of experience without centralized control.

Contribution

It introduces a novel adaptive multicast tree construction method based on a multi-armed bandit framework for open P2P networks.

Findings

DecVi achieves similar quality-of-experience to centralized algorithms.

DecVi provides higher reliability and flexibility in network conditions.

DecVi effectively operates with limited global network knowledge.

Abstract

Video conferencing has become the preferred way of interacting virtually. Current video conferencing applications, like Zoom, Teams or WebEx, are centralized, cloud-based platforms whose performance crucially depends on the proximity of clients to their data centers. Clients from low-income countries are particularly affected as most data centers from major cloud providers are located in economically advanced nations. Centralized conferencing applications also suffer from occasional outages and are embattled by serious privacy violation allegations. In recent years, decentralized video conferencing applications built over p2p networks and incentivized through blockchain are becoming popular. A key characteristic of these networks is their openness: anyone can host a media server on the network and gain reward for providing service. Strong economic incentives combined with lower entry barrier to join the network, makes increasing server coverage to even remote regions of the world. These reasons, however, also lead to a security problem: a server may obfuscate its true location in order to gain an unfair business advantage. In this paper, we consider the problem of multicast tree construction for video conferencing sessions in open p2p conferencing applications. We propose DecVi, a decentralized multicast tree construction protocol that adaptively discovers efficient tree structures based on an exploration-exploitation framework. DecVi is motivated by the combinatorial multi-armed bandit problem and uses a succinct learning model to compute effective actions. Despite operating in a multi-agent setting with each server having only limited knowledge of the global network and without cooperation among servers, experimentally we show DecVi achieves similar quality-of-experience compared to a centralized globally optimal algorithm while achieving higher reliability and flexibility.