Non-blocking k-ary Search Trees

Abstract

This paper presents the first concurrent non-blocking k-ary search tree. Our data structure generalizes the recent non-blocking binary search tree of Ellen et al. [5] to trees in which each internal node has k children. Larger values of k decrease the depth of the tree, but lead to higher contention among processes performing updates to the tree. Our Java implementation uses single-word compare-and-set operations to coordinate updates to the tree. We present experimental results from a 16-core Sun machine with 128 hardware contexts, which show that our implementation achieves higher throughput than the non-blocking skip list of the Java class library and the leading lock-based concurrent search tree of Bronson et al. [3].