FastRank: Practical lightweight tolerance to rational behavior in edge assisted streaming

https://doi.org/10.1016/j.pmcj.2018.02.005Get rights and content

Abstract

Edge-computing is one of the most promising techniques to leverage the excess capacity that exists at users’ premises. Unfortunately, edge-computing may be vulnerable to free-riding, i.e., to nodes that attempt to benefit from the system without providing any service in return. Traditional approaches model free-riders as rational nodes that strive to maximize a utility and apply Game Theory concepts to devise mechanisms that deny any utility gain to nodes that deviate from the protocol. These mechanisms impose significant overhead. This paper proposes a new approach that avoids these overheads by applying concepts of evolutionary game theory. We propose to devise lightweight mechanisms targeted for the optimistic setting where the vast majority of nodes adopts one of a small number of behaviors. More precisely, we assume that most nodes are altruistic or follow non-sophisticated behaviors such as free-riding or white-washing. If a small fraction of nodes follows alternative behaviors, then our lightweight mechanism limits the utility gain of these nodes, making it unlikely that the number of nodes exploiting sophisticated behaviors increases at a fast pace. This allows altruistic nodes to detect their presence in time to switch to more robust mechanisms, before the system reaches a state where the lightweight mechanisms can no longer cope with the existing behaviors. We apply this approach in the context of edge-assisted streaming.

Introduction

Decentralized peer-to-peer systems are a powerful tool to explore the unused capacity at the edges of the network. Unfortunately, it has been observed  [[1], [2], [3]] that a significant fraction of nodes may free-ride, taking advantage of their participation on the system although not providing any useful contribution. Free-riders benefit from the work of the large fraction of altruistic nodes, those that cooperate unconditionally to the system. This not only puts an unfair load on altruistic nodes, but also degrades the performance of the tasks executed at the edge. Mechanisms that can detect free-riders and prevent them from dominating the system operation are therefore extremely relevant.

Existing work  [[4], [5], [6], [7]] has addressed free-riding by adopting a game theoretical approach. In this approach, nodes are assumed to be rational. Rational nodes do not contribute because they aim at maximizing a utility, which decreases with the amount of resources provided to other nodes and increases with the benefits of participating in the system. These works proposed incentive mechanisms that are incentive-compatible by denying any utility gain to nodes that deviate from the protocol. However, incentive-compatible mechanisms have two main drawbacks. First, because rational nodes can deviate from the protocol in an arbitrary fashion, the mechanisms must devise sophisticated incentives, which are generally costly from a system perspective. Second, nodes may not be capable of determining which of two protocols provides a higher utility.

Evolutionary Game Theory (EGT) is an alternative framework to Game Theory that does not assume that nodes are fully rational  [[8], [9]]. Instead, every node n starts by following some “hardwired” protocol. At each point in time, n may change to a different protocol with some probability, either by replicating the behavior of other nodes or by mutating to a new protocol. The replication probability is proportional to the utility difference between protocols. EGT aims to devise protocols that are stable in the sense that, if all nodes are following the protocol and a small fraction of nodes mutates by following another protocol, then this fraction does not tend to grow with time. This models the realistic scenarios where nodes start by following some known protocol (e.g., BitTorrent in file-sharing), but then some of them may discover alternative protocols that they believe to provide a higher utility. Subsequently, other nodes may be influenced by them and also change their behavior.

In this paper, we advocate an adaptive approach inspired in the concepts of EGT to devise efficient and stable incentive mechanisms. The approach comprises at least two different protocols: one is a lightweight protocol that is stable in an optimistic setting where nodes may deviate from the protocol in a small number of ways, whereas the other is a more costly stable protocol for arbitrary settings. While we can find in the literature several (costly) protocols that address arbitrary settings (such as  [[4], [5], [6], [7]]), to the best of our knowledge, there are no examples of lightweight protocols that are both efficient and stable in an optimistic setting. This paper bridges this gap by proposing such a protocol, named FastRank.

Evidence collected from past and current usage of edge-assisted systems, including peer-to-peer file sharing and live-streaming, confirms that in most decentralized systems, most users either volunteers to provide service to others (i.e., are altruistic) or deviate from the protocol by adopting simple free-riding or white-washing behaviors  [[2], [10], [3]]. We argue that the reason why more sophisticated behaviors are not frequently observed is because they require developing an alternative protocol, which most users are not willing or capable to do. Recent experiments have shown that a streaming system cannot provide high reliability if a majority of users deviates  [10]. This motivates us to target the lightweight protocol to an optimistic setting with at most 30% users that free-ride or white-wash. This is intended to capture an initial setting where most users have not yet discovered the advantages of deviating and thus have not mutated into following more beneficial protocols. If a higher fraction of deviating users is expected in the initial setting, then the system designer should require users to follow a more costly and robust protocol from the beginning. To make the system stable to mutations, the lightweight protocol must provide a metric sensible to non-expected behaviors (i.e., different from free-riding and white-washing), so that if altruistic nodes detect that the fraction of nodes deviating in an unexpected way is too large, then they may switch to a more robust protocol. The lightweight protocol must also limit the utility gain of nodes with sophisticated behavior to sustain the growth rate of these nodes until the switch operation is concluded.

To demonstrate the applicability of our approach, we address the problem of edge-assisted streaming. The usefulness of edge-computing to support live streaming has been demonstrated by several large scale real-life deployments, including PPLive  [11] among others  [12]. We propose a lightweight protocol that is stable in a setting with altruists, free-riders, and white-washers. Like most live-streaming implementations, we require nodes to join an overlay network in order to receive and forward packets from a given stream. We consider that free-riders never forward packets, and white-washers strive to circumvent the incentive mechanisms of our protocol by constantly seeking new neighbors from whom they receive the stream. Our overlay maintenance protocol includes incentives for nodes to keep stable symmetric links with a small set of neighbors. As a result, it becomes possible to use simple direct reciprocity mechanisms to deny them the benefit of receiving the stream. We also limit the frequency with which nodes can create new relationships to cope with white-washing behavior. Based on these principles, we propose FastRank, an integrated topology management and peer-monitoring scheme that can effectively and efficiently minimize the impact of free-riders and white-washers. Interestingly, we also show that the approach limits the utility gain of more sophisticated behaviors. Furthermore, we show that, if the fraction of nodes that follow sophisticated behaviors grows to a point where it impacts the reliability, this can be detected by altruistic nodes that can then trigger a system reconfiguration and commute to a more robust protocol.

FastRank has been extensively evaluated. Our experimental results show that FastRank is able to keep altruistic nodes connected to each other, while isolating free-riders, denying them access to the stream, and limiting the frequency with which white-washers create new relationships to the point where they do not gain from this strategy. We also show that these incentives are sufficient to limit the utility gain of nodes adopting a more sophisticated behavior. These receive a limited benefit and, if their fraction is large, their presence can be detected by altruistic nodes in time to switch to a more robust protocol.

The rest of the paper is structured as follows. Section 2 surveys related work. Section 3 presents the model. Sections 4 FastRank, 5 Adaptive framework describe FastRank and how it can be leveraged to build a full adaptive solution. Section 6 reports on our experimental results. Finally, Section 7 concludes the paper.

Section snippets

Related work

There are two main approaches to support live-streaming in large scale systems: one is to build multicast dissemination trees and the other is to rely on gossip protocols. In optimal conditions, i.e. with minimal churn and a negligible number of free-riders, tree based approaches are a natural and efficient structure to spread information as they avoid any redundant message delivery. However, in an uncontrolled environment like the Internet, the performance of tree-based approaches degrades

System model

We focus on the problem of disseminating a stream in a gossip fashion from a source to all the nodes of a peer-to-peer overlay network. The source first partitions the stream into multiple frames and sends them to a subset of nodes. Then nodes cooperate to disseminate the stream by forwarding each received frame to a subset of nodes selected at random. Each node n has an identifier and a view of the system, defined as the subset of identifiers of other nodes to which n forwards frames. The

FastRank

We now describe FastRank, a peer-to-peer streaming service that is stable in a setting with altruists, free-riders, and white-washers. FastRank has three main components: an overlay network construction and maintenance protocol, a localized neighbor ranking mechanism, and a dissemination mechanism. These components cooperate in a synergistic manner to ensure that free-riders are promptly identified and shunned by their neighbors, so that they stop receiving the stream. In particular, the

Adaptive framework

FastRank provides a solution to streaming that is stable in a setting where non-altruistic nodes follow free-riding or white-washing strategies. However, this solution is not stable in general settings, where nodes may adopt other more sophisticated strategies, e.g., forwarding frames with a probability lower than what is specified by the protocol or varying the size of the active view. If a fraction of nodes mutates into following more sophisticated strategies, then this fraction will grow

Evaluation

In this section, we provide an extensive evaluation of FastRank using both simulations and experiments with a prototype on PlanetLab. Simulations were performed using the PeerSim framework  [22]. They consisted in the dissemination of 20 000 frames among 1000 nodes. Frames were injected in the network by the streamer using 7 peers randomly chosen, at a rate of 100 kbps. Results presented in this section are the average of 100 independent runs. We performed an identical evaluation in a PlanetLab

Conclusions

In this paper we have presented FastRank, a peer-to-peer streaming protocol that relies on an overlay network with symmetric links to mitigate the effect of free riders and white-washers in an efficient and effective manner. FastRank includes overlay construction mechanisms that encourage nodes to perform repeated interactions with a small number of nodes and then leverages from this property to implement efficient localized scoring mechanisms that can be used to deny utility gains to

Acknowledgments

This work has been partially supported by Fundação para a Ciência e Tecnologia (FCT) through projects with references PTDC/EEI-SCR/2776/ 2012 (PEPITA), PTDC/EEI-SCR/1741/2014 (Abyss) and UID/CEC/50021/2013.

References (23)

  • AdarE. et al.

    Free riding on gnutella

    First Monday

    (2000)
  • B. Cohen, Incentives build robustness in bittorrent, in: Proceedings of the 1st Workshop on Economics of Peer-to-Peer...
  • HughesD. et al.

    Free riding on gnutella revisited: the bell tolls?

    IEEE Distrib. Syst. Online

    (2005)
  • A. Diarra, S. Mokhtar, P. Aublin, V. Quma, FullReview: Practical accountability in presence of selfish nodes, in:...
  • GuerraouiR. et al.

    Lifting: Lightweight freerider-tracking in gossip

  • H. Li, A. Clement, M. Marchetti, M. Kapritsos, L. Robison, L. Alvisi, M. Dahlin, FlightPath: Obedience vs. choice in...
  • H. Li, A. Clement, E.L. Wong, J. Napper, I. Roy, L. Alvisi, M. Dahlin, Bar gossip, in: Proceedings of the 7th USENIX...
  • HauertC. et al.

    Evolutionary games and population dynamics: maintenance of cooperation in public goods games

    Proc. R. Soc. Lond. Ser. B Biol. Sci.

    (2006)
  • SmithJ.

    Evolution and the Theory of Games

    (1982)
  • J. Oliveira, I. Cunha, E. Miguel, M. Rocha, A. Vieira, S. Campos, Can peer-to-peer live streaming systems coexist with...
  • HuangY. et al.

    Challenges, design and analysis of a large-scale P2P-vod system

    SIGCOMM Comput. Commun. Rev.

    (2008)
  • Cited by (1)

    View full text