Epidemic Protocols (or, Gossip is Good)

10/7/99


Click here to start


Table of Contents

Epidemic Protocols (or, Gossip is Good)

Overview

Data Dissemination

History

How does gossip spread?

State Monotonic Property

Anti-Entropy

How fast does gossip spread?

Probability of Infection?

Intuition: 2 phases

Phase 1: fast growth of infection

Phase 2: fast decline of uninfection

Exponential growth

#rounds distribution

Expected #rounds

Case Study 1: Failure Detection

Informal Properties

Environment

Basic Gossip Protocol

Linear Bandwidth

Model

Failure Caveat

Analysis

Failure Detection Time

Quality of Detection

Effect of Failed Members

Effect of Message Loss

Case Study 2: Bimodal Multicast

PPT Slide

PPT Slide

PPT Slide

An aside: Push vs. Pull

PPT Slide

PPT Slide

Example

Delivery? Garbage Collection?

Throughput with perturbed processes

But, gossip doesn’t scale...

High load on routers

Problems with partitions

Idea: add locality to gossip

Domains

Multi-level Gossip Protocol

Better properties

Two-level hierarchy

Two-level cost

No longer logarithmic...

Problems

New idea

Case Study 3: Astrolabe

Example HMIB

HMIB tree

HMIB node/table

Example: digital library

Applications

Other applications

Implementation

Membership issues

Failure detection

Member detection / Partition resolver

Condensation Functions

A few word on security

Related Literature

Author: Robbert van Renesse

Email: rvr@cs.cornell.edu

Home Page: http://www.cs.cornell.edu/home/rvr