Constant RMR solutions to reader writer synchronization

Abstract

We study Reader-Writer Exclusion [1], a well-known variant of the Mutual Exclusion problem [2] where processes are divided into two classes - readers and writers - and multiple readers can be in the Critical Section (CS) at the same time, although no process may be in the CS at the same time as a writer. Since readers don't conflict with each other, they should not obstruct each other. Specifically, the concurrent entering property must be satisfied: if all writers are in the Remainder section, each reader should be able to enter the CS in a bounded number of its own steps. Three versions of the Reader-Writer Exclusion problem are commonly studied - one where writers have priority over readers, another where readers have priority, and the last where neither class has priority over the other and no process may starve. To ensure high performance on Cache-Coherent (CC) and Distributed Shared Memory (DSM) multiprocessors, algorithms should be designed to generate as few remote memory references (RMRs) as possible. It would be ideal to achieve constant RMR complexity, i.e., the worst case number of RMRs that a process generates in order to enter and exit the CS once is a constant, independent of the number of processes. Constant RMR complexity algorithms have existed for Mutual Exclusion for two decades [3, 4], but none exists for Reader-Writer Exclusion. Danek and Hadzilacos' lower bound proof implies that it is impossible to achieve sublinear RMR complexity for DSM machines [5]. For CC machines, the best existing bound, also due to Danek and Hadzilacos [5], is O(log n), where n is the number of processes. In this work, we present the first constant RMR complexity algorithms for all three versions of the Reader-Writer Exclusion problem (for CC machines).