On the robustness of (semi) fast quorum-based implementations of atomic shared memory
Chryssis Georgiou, Nicolas Nicolaou, Alexander A. Shvartsman · 2008
Atomic (linearizable) read/write memory is a fundamental abstractions in distributed computing. Following a seminal implementation of atomic memory of Attiya et al. [6], a folklore belief developed that in messaging-passing atomic memory implementations "reads must write." However, work by Dutta et al. [4] established that if the number of readers R is constrained with respect to the number of replicas S and the maximum number of crash-failures t so that R < S/t - 2, then single communication round-trip reads are possible. Such an implementation given in [4] is called fast. Subsequently, Georgiou et al. [3] relaxed the constraint in [4], and proposed semifast implementations with unbounded number of readers, where under realistic conditions most reads need only a single communication round-trip to complete. Their approach groups collections of readers into virtual nodes. Semifast behavior of their algorithm is preserved as long as the number of virtual nodes V is constrained by V < S/t - 2.