AGREEMENT IN NETWORKS OF Sample Clauses

AGREEMENT IN NETWORKS OF. GROUPS With data centers spread across the globe and dis- tributed systems that span the Internet, the problem of en- suring consistency in distributed environments has gained new dimensions. The busyness model of on-line retailers, such as Xxxxxx.xxx, and social networking sites, such as Facebook and Orkut, mandate that the user data, such as profiles and shopping carts, be available to their millions of users at nearby dataceters for a better user experience. Data must also be available at other locations to support mobility. In some cases, although clients may access their data on any data center, updates are performed in a sin- gle location and only later propagated to the other loca- tions. While simpler to implement, this strategy incurs in higher latency between an update and the availability of the written data at the closest datacenter; allowing updates everywhere, instead, is preferable. Implementing such an update-everywhere database replication may be done atop of agreement protocols (e.g., Xxxxxx et al.[30]) but, to be effective, such agreement protocols must be efficient and resilient in this networking scenario which we refer to as networks of groups. { } Networks of groups may be abstracted as groups of agents and, typically, are characterized by large differ- ences in terms of latency of communication between two agents: while agents within the same group exchange messages through low-latency communication channels, those in different groups may experience latencies that are orders of magnitude higher. More specifically, we con- sider networks abstracted by Figure 2, in which agents are organized in a set of m subsets or groups Γ = G1, . . . , Gm . Although groups may be seen as data centers, the same abstraction also works for smaller se- tups as, for example, the internals of a single data center, where groups are racks of nodes. Besides the sets G1, . . . , Gm of agents that effectively propose and learn the agreed values, Figure 2 depicts the acceptors in the system as another set, A. To ensure fault tolerance, A should be composed of agents executing in the same physical locations as the agents in Γ. Hence, we consider that the cost of exchanging information between any group and A and between the elements of A is as expensive as between two agents in different groups.