From Epidemics to Distributed Computing

— Epidemic algorithms have been recently recognized as robust and scalable means to disseminate information in large-scale settings. Information is disseminated reliably in a distributed system the same way an epidemic would be propagated throughout a group of individuals: each process of the system chooses random peers to whom it relays the information it has received. The underlying peer-to-peer communication paradigm is the key to the scalability of the dissemination scheme. Epidemic algorithms have been studied theoretically and their analysis is built on sound mathematical foundations. Although promising, their general applicability to large scale distributed systems has yet to go through addressing many issues. These constitute an exciting research agenda. The traditional client-server computing model is not adequate to address reliability and scalability properties. Even when servers are replicated for fault-tolerance, the synchronisation mechanism needed to preserve replica consistency is a major source of overhead. The peer-to-peer computing model represents a radically different and a plausible alternative for many large scale applications in Internet-wide settings. In this model, resources are shared between end-user processes themselves in a peer style, meaning that every such process potentially acts both as a client and a server. Central points of failures disappear as well as the associated performance bottlenecks. Scalability is achieved because the load, whether it consists in forwarding messages or storing data for example, is balanced between all processes of the system. In addition, each process requires only local knowledge of the state of the system. Application-level multicast is one of the distributed applications which may benefit from such a communication paradigm to scale to large systems. The design of scalable peer-to-peer application-level multicast protocols is not straightforward and represents an active area of research. This is one of the challenges that epidemic information dissemination algorithms take up [2]. EPIDEMIC INFORMATION DISSEMINATION Epidemic algorithms have recently gained popularity as a robust and scalable way of propagating information in distributed systems [1]. A process that wishes to disseminate a new piece of information to the system does not send it to a server, or a cluster of servers, in charge of forwarding it, but rather to a set of other peer processes, chosen at random. In turn, each of these processes does the same, and also forwards the information to randomly selected processes, and so forth. The principle underlying this information dissemination technique mimics the spread of epidemics, and we talk about epidemic …