Concurrent Linearizable Nearest Neighbour Search in LockFree-kD-tree

Abstract

The Nearest neighbour search (NNS) is a fundamental problem in many application domains dealing with multidimensional data. In a concurrent setting, where dynamic modifications are allowed, a linearizable implementation of NNS is highly desirable. This paper introduces the LockFree-kD-tree (LFkD-tree): a lock-free concurrent kD-tree, which implements an abstract data type (ADT) that provides the operations Add, Remove, Contains, and NNS. Our implementation is linearizable. The operations in the LFkD-tree use single-word read and compare-and-swap (CAS) atomic primitives, which are readily supported on available multi-core processors. We experimentally evaluate the LFkD-tree using several benchmarks comprising real-world and synthetic datasets. The experiments show that the presented design is scalable and achieves significant speed-up compared to the implementations of an existing sequential kD-tree and a recently proposed multidimensional indexing structure, PH-tree.