an approximate matching

peer-to-peer overlay

Main Download Approach Papers Support Press People


Cubit is a scalable peer-to-peer system that can efficiently find the k closest data items to any search key. The central insight behind Cubit is to create a keyword metric space that captures the relative similarity of keywords, to assign portions of this space to nodes in a light-weight overlay and to resolve queries by efficiently routing them through this space. The system comprises a protocol for object and node assignment, a gossip-based protocol for maintaining the overlay, and a routing protocol to efficiently route queries.

Keyword Space

Keyword space

Figure 1: The edit-distance between keywords. The five keywords create a keyword space that cannot be accurately embedded into a plane.

A keyword is any word that appears in the title of an object stored in Cubit. In order to fully specify the problem of approximate string matching, we need to choose a notion of distance between two keywords, or more generally between two text strings. Such distance should correspond to our intuition on which strings are similar and which strings are very different.

Cubit uses the most common notion of distance on strings, the edit-distance. It is equal to the minimum number of insertions, deletions and substitutions needed to transform one string to another. The keywords then lie in the keyword space, illustrated in Figure 1, a metric space induced on keywords by the edit-distance.

Node ID Assignment

Cubit nodes are distributed in the same space as keywords. Each node in Cubit is assigned a unique string ID chosen from the set of keywords associated with previously inserted objects in the system. Specifically, at join time each node independently selects a random keyword, ensuring uniqueness by detecting ID collisions. The ID of a node determines its "position" in the keyword space. This position determines how a given node is used in Cubit. First, each Cubit node is responsible for storing the set of keywords for which it is the closest node. Second, Cubit implements a distributed protocol which navigates through nodes in the keyword space, gradually zooming in on a neighborhood of a given (possibly misspelled) keyword, and thus locates nodes that store possible matches.


Multi-resolution rings

Figure 2: In this example, the solid circles represent peers in node A's peer-set, the empty circles represent other peers, and the squares represent object keywords in the system. The shaded region depicts the sub-space that is closer to A than any other node.

The navigation protocol is the core component of Cubit. To support this protocol, Cubit creates and maintains a multi-resolution overlay network on nodes such that each node has several peers at every distance from itself; the peers at a given distance are chosen to maximize the coverage of that region. This is illustrated in Figure 2. Such overlay design is inspired by the small-world construction in which a grid is augmented by a sparse set of randomly chosen edges, with roughly the same number of edges for each distance scale. In the resulting graph a simple greedy routing algorithm (which on each step minimizes the distance to target) succeeds in finding short routes to any given target.

The desired property of the search protocol is to obtain the k closest objects to the set of keywords, as measured by the phrase distance metric. For each keyword in the search phrase, the protocol obtains the k closest objects from each node which meets the following edit distance criterion: its ID is within an edit-distance of q from the keyword, where q is the product of the keyword length and the expected number of perturbations per character (which is a parameter in the system). The protocol selects m closest nodes if fewer than m nodes meet edit-distance criterion, where m is called the search fan-out. The keyword search protocol is illustrated in Figure 3.

Search protocol

Figure 3: The Cubit search protocol operates iteratively to collect more and more information of the target region. In this example, x is the location of the search term in the keyword space, the solid circles are node A's peers, empty circles are additional nodes in the space, and the circle around x are all nodes within edit-distance q of x. Node A first finds the m = 2 closest nodes to x from its peer-set, and request their m closest nodes. In this example, two new closer nodes are discovered and subsequently sent the same query. The protocol terminates when all nodes within the circle around x, or the m closest nodes have been discovered. These nodes are queried for their closest objects to x.

Load Balancing

Load balance

Figure 4: In this example, the keyword "love" is closest to node A and is generating a high degree of load. Node A creates a virtual node centered around the keyword love, which includes its leaf set and all objects in the region within p edit-distance from love. This virtual node is sent to A's nearest neighbors. Queries that arrive at these neighbors for keywords within an edit-distance p of love can be answered without node A.

Since search terms tend to follow a Zipf distribution, the resulting skewed load distribution can lead to excess routing load on nodes within the vicinity of popular keywords.

In Cubit, if the load generated by queries for a popular keyword w overwhelms the available resources of node i, the node can send an off-loading request to its closest neighbors requesting them to create a synthetic node located at w. Nodes receiving such a request create a synthetic node at w whose IP address and port correspond to their own, thus enabling queries for that portion of the keyword space to be terminated at any one of the moff neighbors. The original requester is then tasked with keeping the virtual nodes updated with changes to objects in the off-loaded region as well as changes to its leaf-set. This off-loading operation disperses hot-spots in keyword popularity without requiring global information or coordination. Figure 4 illustrates the protocol.

Cubit Project

Computer Science Department
Cornell University