Gossip protocol Totally Explained

Everything about Gossip Protocol totally explained

A gossip protocol is a style of computer-to-computer communication protocol inspired by the form of gossip seen in social networks. Modern distributed systems often use gossip protocols to solve problems that might be difficult to solve in other ways, either because the underlying network has an inconvenient structure, is extremely large, or because gossip solutions are sometimes the most efficient ones available.
The term epidemic protocol is sometimes used as a synonym for a gossip protocol, because gossip spreads information in a manner similar to the spread of a virus in a biological community.

Gossip Communication

We use the term protocol to describe a structured interchange of messages between agents running on a set of computers in a network, designed to achieve some goal. A gossip protocol is a protocol designed to mimic the way that information spreads when people gossip about some fact.
   For example, imagine a group of office workers. Ted comments to Sally that he believes Fred is starting to dye his mustache. Sally tells Jill, while Ted repeats the news to Sam. Notice that as people move about and share the news, the number of individuals who know about the rumor doubles "round by round". Of course this is just an approximation; it assumes that employees run into one-another at a constant rate, and doesn't capture the issue of gossiping twice to the same person (perhaps Ted tries to tell his story to Mark, only to find that Mark already heard it from Jill). However, the approximation is a reasonably good one. In computer systems, we typically implement protocols such as this with some form of random "peer selection": at some wired-in frequency, each machine picks some other machine at random and shares any hot rumors.
   The power of gossip is that information spreads so robustly. For example, even if Jill had trouble understanding Sally (perhaps she was whispering), she'll probably run into someone else soon and can always learn the news that way.
   Expressing these ideas in more technical terms, a gossip protocol is one that satisfies the following conditions:

The core of the protocol involves periodic, pairwise, inter-process interactions
The information exchanged during these interactions is of (small) bounded size
When agents interact, the state of one or both changes in a way that reflects the state of the other. For example, if A pings B just to measure the round-trip time for messages from A to B and back, it isn’t a gossip interaction.
Reliable communication isn't assumed
The frequency of the interactions is low compared to typical message latencies, so that the protocol costs are negligible
There is some form of randomness in the peer selection. Peer selection might occur within the full node set, or might be performed in a smaller set of neighbors.

It is useful to distinguish three prevailing styles of gossip protocol.

Dissemination (rumor-mongering) protocols. These use gossip to spread information; they basically work by flooding agents in the network, but in a manner that produces bounded worst-case loads: ť a) Event dissemination protocols use gossip to carry out multicasts. They report events, but the events don’t actually trigger the gossip (since gossip runs periodically). One concern here's the potentially high latency from when the event occurs until it's delivered.

b) Background data dissemination protocols continuously gossip about information associated with the participating nodes. Typically, propagation latency isn’t a concern, perhaps because the information in question changes slowly or there's no significant penalty for acting upon slightly stale data.

Anti-entropy protocols for repairing replicated data, which operate by comparing replicas and reconciling differences.

Protocols that compute aggregates. These compute a network-wide aggregate by sampling information at the nodes in the network and combining the values to arrive at a system-wide value – the largest value for some measurement nodes are making, smallest, etc. The key requirement is that the aggregate must be computable by fixed-size pairwise information exchanges; these typically terminate after a number of rounds of information exchange logarithmic in the system size, by which time an all-to-all information flow pattern will have been established. As a side effect of aggregation, it's possible to solve other kinds of problems using gossip; for example, there are gossip protocols that can arrange the nodes in a gossip overlay into a list sorted by node-id (or some other attribute) in logarithmic time using aggregation-style information exchanges. Similarly, there are gossip algorithms that arrange nodes into a tree, and then compute aggregates such as "sum" or "count" by gossiping in a pattern biased to match the tree structure. This is a rather inclusive definition. Many protocols that predate the earliest use of the term “gossip” fall within this definition. For example, Internet routing protocols often use gossip-like information exchanges.
   It is interesting to realize that a gossip substrate can easily implement a standard routed network: nodes could “gossip” about traditional point-to-point messages, effectively tunneling normal traffic through the gossip layer. Bandwidth permitting, this implies that a gossip system can potentially support any classic protocol or implement any classical distributed service. Nonetheless, when we talk of gossip protocols, we rarely intend such a broadly inclusive interpretation. More typically we've in mind protocols that run in a regular, periodic, relatively lazy, symmetric and decentralized manner; the high degree of symmetry among nodes is particularly characteristic. Thus, to give a simple example, while one could run a 2-phase commit protocol over a gossip substrate, piggybacking the messages on gossip traffic, doing so would be at odds with the spirit, if not the wording, of the definition.
   Frequently, the most useful gossip protocols turn out to be those with exponentially rapid convergence towards a state that “emerges” with probability 1.0. For example, a classic distributed computing problem involves building some form of tree whose inner nodes are the nodes in a network, and whose edges represent links between computers (for routing, as a dissemination overlay, etc). Not all tree-building protocols are gossip protocols (for example, spanning tree constructions in which a leader initiates a flood), but gossip offers a good decentralized solution that can be useful in many situations.
   The term convergently consistent is sometimes used to describe protocols that achieve exponentially rapid spread of information. Notice that for this purpose, a protocol must propagate any new information to all nodes that will be affected by the information within time logarithmic in the size of the system (the “mixing time” must be logarithmic in system size).

Examples

Suppose that we want to find the object that most closely matches some search pattern, within a network of unknown size, but where the computers are linked to one-another and where each machine is running a small agent program that implements a gossip protocol.

To start the search, a user would ask the local agent to begin to gossip about the search string. (We're assuming that agents either start with a known list of peers, or retrieve this information from some kind of a shared web site.)

Periodically, at some rate (let's say ten times per second, for simplicity), each agent picks some other agent at random, and gossips with it. Search strings known to A will now also be known to B, and vice versa. In the next "round" of gossip A and B will pick additional random peers, maybe C and D. This round-by-round doubling phenomenon makes the protocol very robust, even if some messages get lost, or some of the selected peers are the same or already know about the search string.

On receipt of a search string for the first time, each agent checks its local machine for matching documents.

Agents also gossip about the best match, to date. Thus, if A gossips with B, after the interaction, A will know of the best matches known to B, and vice versa. Best matches will "spread" through the network. If the messages might get large (for example, if many searches are active all at the same time), a size limit should be introduced. Also, searches should "age out" of the network.
   It should be easy to see that within logarithmic time in the size of the network (the number of agents), any new search string will have reached all agents. Within an additional delay of the same approximate length, every agent will learn where the best match can be found. In particular, the agent that started the search will have found the best match.
   For example, in a network with 25,000 machines, we can find the best match after about 30 rounds of gossip: 15 to spread the search string and 15 more to discover the best match. A gossip exchange could occur as often as once every tenth of a second without imposing undue load, hence this form of network search could search a big data center in about 3 seconds.
   In this scenario, searches might automatically age out of the network after, say, 10 seconds. By then, the initiator knows the answer and there's no point in further gossip about that search.
   Gossip protocols have also been proposed for such tasks as maintaining databases or other kinds of files in consistent states, counting the number of nodes in a network of unknown size, spreading news robustly, organizing nodes according to some structuring policy, building so-called overlay networks, computing aggregates, sorting the nodes in a network, electing leaders, etc.

Epidemic Algorithms

Gossip protocols can be used to propagate information in a manner rather similar to the way that a viral infection spreads in a biological population. Indeed, the mathematics of epidemics are often used to model the mathematics of gossip communication. The term epidemic algorithm is sometimes employed when describing a software system in which this kind of gossip-based information propagation is employed.

Biased Gossip

Above, we described a purely random peer-selection scheme for gossip: when agent A decides to run a gossip round, it picks some peer B uniformly and at random within the network as a whole (or launches a message on a random walk that will terminate at a random agent). More commonly, gossip algorithms are designed so that agents interact mostly with nearby agents, and only sometimes with agents that are far away (in terms of network delay). These biased gossip protocols need to ensure a sufficient degree of connectivity to avoid the risk of complete disconnection of one side of a network from the other, but if care is taken, can be faster and more efficient than protocols that are purely random. Moreover, as a purely practical question, it's much easier to maintain lists of peers in ways that might be somewhat biased.

Further Information

Get more info on 'Gossip Protocol'.

External Link Exchanges

Do you know how hard it is to get a link from a large encyclopaedia? Well we're different and will prove it. To get a link from us just add the following HTML to your site on a relevant page:

<a href="http://gossip_protocol.totallyexplained.com">Gossip protocol Totally Explained</a>

Then simply click through this link from your web page. Our crawlers will verify your link, extract the title of your web page and instantly add a link back to it. If you like you can remove the words Totally Explained and embed the link in article text.
As long as your link remains in place, we'll keep our link to you right here. Please play fair - our crawlers are watching. Your site must be closely related to this one's topic. Any kind of spamming, dubious practises or removing the link will result in your link from us being dropped and, potentially, your whole site being banned.