Robust emulation of shared memoryusing dynamic quorum - acknowledged

This paper presents robust emulation of multi-writer/multi-reader registers in message-passing systems using dynamic quorum conngurations. In addition to processor and link failures, this emulation tolerates changes in quorum conngurations, i.e., on-line replacements of one quorum system consisting of read and write quorums with another such system. This work extends the results of Attiya, Bar-Noy and Dolev 1] who showed how to emulate single-writer/multi-reader registers robustly in message-passing systems using majorities. The emulation in this paper is speciied using a modular two-layer architecture. The lower layer uses unreliable broadcast to disseminate a request from the higher layer to a set of processors, and then to collect responses from a subset of the processors. The subset can be speciied by a predicate or by using a quorum system. The lower layer then computes a function on the collected responses and returns the result to the higher layer. The broadcast can take advantage of hardware-assisted broadcast as we do not assume that the broadcast is reliable or that it has fifo, causal or atomic properties. The higher layer algorithm emulates robust multi-writer/multi-reader registers where quorum conngurations are used to ensure that the registers are atomic. A unique feature of the read/write service is that it implements dynamically changing quorum conngurations. The service includes two interfaces, a functional interface for reads and writes, and a management interface for reconnguration. The processor designated as the reconngurer executes requests that replace the current quorum connguration with the new connguration. The combination of the higher and lower layers allows essentially unlimited concurrency and does not involve locks. Waiting can occur only (a) due to processor or link failures that disconnect at least one processor in each read and write quorum of the speciied conngurations, or (b) when frequent reconngurations interfere with reads/writes and cause them to contribute to reconngurations. However , as soon as reconngurations stop, and as long as for each lower level request specifying a set of read or write quorums there exists a single quorum of active and connected processors, then reads and writes complete without waiting. All of this is transparent to the clients of the service. The algorithms are speciied here in terms of I/O automata 8, 9], and their correctness is proven using invariants and partial-order-based methods. It is shown that the algorithm is correct, and that it implements atomic replicated read/write objects.