Concurrent updates with RCU: search tree as an example

Abstract

Read copy update (RCU) is a novel synchronization mechanism, in which the burden of synchronization falls completely on the updaters, by having them wait for all pre-existing readers to finish their read-side critical section. This paper presents citrus, a concurrent binary search tree (BST) with a wait-free Contains operation, using RCU synchronization and fine-grained locking for synchronization among updaters. This is the first RCU-based data structure that allows concurrent updaters. While there are methodologies for using RCU to coordinate between readers and updaters, they do not address the issue of coordination among updaters, and indeed, all existing RCU-based data structures rely on coarse-grained synchronization between updaters. Experimental evaluation shows that \citrus beats previous RCU-based search trees, even under mild update contention, and compares well with the best-known concurrent dictionaries.