GeoPaxos+: Practical Geographical State Machine Replication

Abstract

In some online services, the geographical location of a client tends to determine the data accessed by the client's requests. Geographical locality holds, for example, in location-based services, tracking systems, and social networking services. State machine replication protocols can use geographical locality to optimize performance by ordering requests efficiently. In order to be effective, though, two requirements must be fulfilled. First, protocols must identify the data accessed by a request before the request is executed. Second, protocols must determine which parts of the service state are accessed where and with what probability. The paper presents a geographical state machine replication protocol that meets both requirements. We illustrate the use of our protocol by developing a geographically replicated B+ Tree service. We fully implemented the B+ Tree service and show experimentally that it outperforms implementations based on classic (i.e., Paxos) and recent (i.e., EPaxos) general-purpose replication protocols by a large margin.