An Analytical Model of Information Dissemination for a Gossip-Based Protocol

Abstract

We develop an analytical model of information dissemination for a gossiping protocol that combines both pull and push approaches. With this model we analyse how fast an item is replicated through a network, and how fast the item spreads in the network, and how fast the item covers the network. We also determine the optimal size of the exchange buffer, to obtain fast replication. Our results are confirmed by large-scale simulation experiments.

On the uniformity of peer sampling based on view shuffling

“…However, these analyses aims to provide some theoretical outcomes on a specific characteristic of these protocols as convergence speed of dissemination protocols, by defining precise lower and upper bounds of the mixing time, degree balancing, etc. Nevertheless, in these works, except for [2], authors do not consider the local view as the information to analyze. In their works, the network is modeled as a probabilistic matrix, which represents the meeting probability of any pair of peers, and this matrix is used as a building block of their analyses.…”

“…On the other hand, Bakhshi et al propose in [2] an analytical model of shuffling protocols. However, as they are interested in the characteristics of the dissemination of data items, they investigate two other properties than ours that characterized these protocols: (i) the number of replicas of an item in the network, and (ii) the coverage achieved by an item over time.…”

“…Several contributions provided some fully theoretical analysis of gossip-based protocols as [2,[15][16][17]8]. However, these analyses aims to provide some theoretical outcomes on a specific characteristic of these protocols as convergence speed of dissemination protocols, by defining precise lower and upper bounds of the mixing time, degree balancing, etc.…”

“…Experimental approaches [2,15,12,14,23] point out that, in the design of a gossip-based protocol, l has to be set to the half of c in order to obtain the highest efficiency in terms of convergence speed. This conjecture can be intuitively shown as sketched below.…”

“…Each node i manages a local partial view of the system, denoted V i of size c ≤ n, about all the other nodes in the system, including itself. 2 In other word, each node owns a small set of connections (aka neighbors) corresponding to its vision of the overall system. 1 A simple safe condition for keeping the sample returned by a node uniform is to ''refresh'' the returned ID.…”