Consistent distributed memory services: Resilience and effciency

Reading, 'Riting, and 'Rithmetic, the three R's underlying much of human intellectual activity, not surprisingly, also stand as a venerable foundation of modern computing technology. Indeed, both the Turing machine and von Neumann machine models operate by reading, writing, and computing, and all practical uniprocessor implementations are based on performing activities structured in terms of the three R's. With the advance of networking technology, communication became an additional major systemic activity. However, at a high level of abstraction, it is apparently still more natural to think in terms of reading, writing, and computing. While it is hard to imagine distributed systems-such as those implementing the World-Wide Web-without communication, we often imagine browser-based applications that operate by retrieving (i.e., reading) data, performing computation, and storing (i.e., writing) the results. In this article, we deal with the storage of shared readable and writable data in distributed systems that are subject to perturbations in the underlying distributed platforms composed of computers and networks that interconnect them. The perturbations may include permanent failures (or crashes) of individual computers, transient failures, and delays in the communication medium. The focus of this paper is on the implementations of distributed atomic memory services. Atomicity is a venerable notion of consistency, introduced in 1979 by Lamport [35]. To this day atomicity remains the most natural type of consistency because it provides an illusion of equivalence with the serial object type that software designers expect. We define the overall setting, models of computation, definition of atomic consistency, and measures of e ciency. We then present algorithms for single-writer settings in the static models. Then we move to presenting algorithms for multi-writer settings. For both static settings we discuss design issues, correctness, e ciency, and trade-o s. Lastly we survey the implementation issues in dynamic settings, where the universe of participants may completely change over time. Here the expectation is that solutions are found by integrating static algorithms with a reconfiguration framework so that during periods of relative stability one benefits from the e ciency of static algorithms, and where during the more turbulent times performance degrades gracefully when reconfigurations are needed. We describe the most important approaches and provide examples.