Maxime Monod

Boosting Gossip for Live Streaming

Gossip protocols are considered very effective to disseminate information in a large scale dynamic distributed system. Their inherent simplicity makes them easy to implement and deploy. However, whereas their probabilistic guarantees are often enough to disseminate data in the context of low- bandwidth applications, they typically do not suffice for high-bandwidth content dissemination: missing 1% is unacceptable for live streaming. In this paper, we show how the combination of two simple mechanisms copes with this seemingly inherent deficiency of gossip: (i) codec, an erasure coding scheme, and (ii) claim², a content- request scheme that leverages gossip duplication to diversify the retransmission sources of missing information. We show how these mechanisms can effectively complement each other in a new gossip protocol, gossip++, which retains the simplicity of deployment of plain gossip. In a realistic setting with an average bandwidth capability (800kbps) close to the stream rate (680kbps) and 1% message loss, plain gossip can provide at most 99% of the stream. Using gossip++, on the other hand, all nodes can view a perfectly clear stream.

Decentralized polling with respectable participants

We consider the polling problem in a social network: participants express support for a given option and expect an outcome reflecting the opinion of the majority. Individuals in a social network care about their reputation: they do not want their vote to be disclosed or any potential misbehavior to be publicly exposed. We exploit this social aspect of users to model dishonest behavior, and show that a simple secret sharing scheme, combined with lightweight verification procedures, enables private and accurate polling without requiring any central authority or cryptography. We present DPol, a simple and scalable distributed polling protocol in which misbehaving nodes are exposed with positive probability and in which the probability of honest participants having their privacy violated is traded off against the impact of dishonest participants on the accuracy of the polling result. The trade-off is captured by a generic parameter of the protocol, an integer k called the privacy parameter. In a system of N nodes with B dishonest participants, the probability of disclosing a participants vote is bounded by (B/N)^k^+^1, whereas the impact on the score of each polling option is at most (3k+2)B, with high probability when dishonest users are a minority (i.e., B

Stretching gossip with live streaming

Gossip-based information dissemination protocols are considered easy to deploy, scalable and resilient to network dynamics. They are also considered highly flexible, namely tunable at will to increase their robustness and adapt to churn. So far however, they have mainly been evaluated through simulation, very often assuming ideal settings. Instead, in this paper, we report on an extensive study of gossip protocols, deployed on a 230 Planetlab node testbed, in the context of a challenging video streaming application in environments with constrained bandwidths. More precisely, we assess the impact of varying the well known knobs of gossip, fanout and refresh rate, in various upload-bandwidth distributions and churn. Our results show that in such challenging contexts, the performance of gossip protocols may be hampered by high fanout values. We also show that the more proactive a gossip protocol, the better it copes with churn. For instance, when 20% of the nodes simultaneously crash, 70% of the remaining nodes do not suffer any loss in stream quality, while the others only experience a performance decrease for an average of 5 seconds around the churn event.

A generic theoretical framework for modeling gossip-based algorithms

During the past 30 years of the Internet revolution, the Internet has become a major force of change with an enormous effect on civilization. Consequently, computer networks have evolved into more complex system and become virtually ubiquitous. This in turn, has given raise to a growing demand for scalable and reliable computer system architectures. Thus far, there has been enormous effort by the research community to introduce decentralized, simple, and scalable distributed systems to solve a wide range of problems. In this paper we explore one promising solution, which was initially inspired by mathematical models that investigate two everyday life phenomena, epidemics and gossip, which we used interchangeably throughout the paper. During the last century, mathematicians developed models to predict the rate of diseases spread, namely epidemics, using differential equations. In addition, researches developed discrete mathematics models to predict what we already know; rumors spread fast, namely gossip. It was thus natural to harness these models in order to design distributed systems that mimic the basic behavior of such fast spreading everyday life paradigms.

Decentralized Polling with Respectable Participants

We consider the polling problem in a social network where participants care about their reputation: they do not want their vote to be disclosed nor their misbehaving, if any, to be publicly exposed. Assuming this reputation concern, we show that a simple secret sharing scheme, combined with verification procedures, can efficiently enable polling without the need for any central authority or heavyweight cryptography. More specifically, we present DPol, a simple and scalable distributed polling protocol where misbehaving nodes are exposed with a non-zero probability and the probability of dishonest participants violating privacy is balanced with their impact on the accuracy of the polling result. The trade-off is captured by a generic parameter of the protocol, an integer k we call the privacy parameter , so that in a system of N nodes with

Pervasive Computing with Frugal Objects

B dishonest participants, the probability of disclosing a participants vote is bounded by (B /N ) k + 1, whereas the impact on the polling result is bounded by (6k + 2) B . We report on the deployment of DPolover 400 PlanetLab nodes. The polling result suffers a relative error of less than 10% in the face of message losses, crashes and asynchrony inherent in PlanetLab. In the presence of dishonest nodes, our experiments show that the impact on the polling result is (4k + 1) B on average, consistently lower that the theoretical bound of (6k + 2) B .

Towards Fair Event Dissemination

This paper presents a computing model for resource-limited mobile devices that might be ubiquitously deployed in private and business environments. The model integrates a strongly-typed event-based communication paradigm with abstractions for frugal control, assuming a small footprint runtime. With our model, an application consists of a set of distributed reactive objects, called frugal objects (FROBs), that communicate through typed events and dynamically adapt their behavior according to notifications about changes in resource availability. FROBs have a logical time-slicing execution pattern that helps monitor resource consuming tasks and determine resource profiles in terms of CPU, memory, battery and bandwidth.

Tracking freeriders in gossip-based content dissemination systems

Event dissemination in large scale dynamic systems is typically claimed to be best achieved using decentralized peer-to-peer architectures. The rationale is to have every participant in the system act both as a client (information consumer) and as a server (information dissemination enabler), thus, precluding specific brokers which would prevent scalability and fault-tolerance. We argue that, for such decentralized architectures to be really meaningful, participants should serve the system as much as they benefit from it. That is, the system should be fair in the sense that the extend to which a participant acts as a server should depend on the extend to which it has the opportunity to act as a client. This is particularly crucial in selective information dissemination schemes where clients are not all interested in the same information. In this position paper, we discuss what a notion of fairness could look like, explain why current architectures are not fair, and raise several challenges towards achieving fairness.

Brief announcement: towards secured distributed polling in social networks

Gossip-based protocols have proven very efficient for disseminating high-bandwidth content such as video streams in a peer-to-peer fashion. However, for the protocols to work, nodes are required to collaborate by devoting a fraction of their upload bandwidth, a scarce resource for some of them, to forward the content they receive to other nodes. Consequently, such protocols suffer from freeriding, a common phenomenon on the Internet, which consists in selfishly benefiting from the system without contributing its fair share. Due to the dynamic nature and the inherent randomness of gossip protocols and to the high scalability requirements of video streaming systems, detecting freeriders is a difficult challenge. This paper presents LiFTinG, the first protocol for detecting freeriders, including colluding ones, in gossip-based content dissemination systems with asymmetric data exchanges. In addition, LiFTinG is still able to detect freeriders when network coding, a widely used technique to improve the efficiency of content dissemination, is used. LiFTinG relies on nodes to track abnormal behavior by cross-checking the history of their previous interactions and exploits the fact that nodes pick neighbors at random to prevent colluding nodes from mutually covering up their bad actions. We present a methodology for setting the parameters of LiFTinG to their optimal value, based on a theoretical analysis and we quantify theoretically the performance of LiFTinG. We derive, based on simulations, the optimal strategy of freeriders by taking into account, through a utility function, the benefit of freeriding and the probability of being detected. In addition to these simulations, we report on the deployment of LiFTinG on PlanetLab. In a 300-node system, where a stream of 674kbps is broadcasted, LiFTinG incurs a maximum overhead of only 8% and provides good detection results: For instance, with 10% of freeriders decreasing their contribution by up to 30%, LiFTinG detects 86% of the freeriders after only 30s and wrongfully expels only a few honest nodes (most of them actually being buggy).

The Weight-Watcher Service and its Lightweight Implementation

Social networks are growing exponentially, and one of the most celebrated examples, Facebook, currently boasts more than 250 million users. A particularly important task in such networks is polling, such as the recent one about the terms of service of Facebook [1]. A defining characteristic of such networks is the one to one mapping between social network identities and real ones (as opposed to virtual world platforms such as SecondLife). Participants in social networks are respectable, that is they do care about their reputation: information related to a user is considered to reflect intimately on the associated real person. We claim that leveraging the fact that users of social networks are concerned over their reputation, we can achieve polling in a distributed manner in the presence of malicious users without the use of heavyweight cryptography.