Radix tree for binary sequences on GPU

Abstract

In this paper, we present radix tree index structure (R-Trie) able to perform lookup over a set of keys of arbitrary length optimized for GPU processors. We present a fully parallel SIMD organized creation and search strategies. The R-Trie supports configurable bit stride for each level and nodes statistics for optimization purposes. We evaluate the performance using two search strategies and Longest Prefix Match (LPM) problem for computer networks. Unlike dedicated LPM algorithms we do not incorporate knowledge about the data or the network masks statistics into the tree construction or algorithm behavior. Our solution may be used in general purpose indexing structures where a batch search of massive number of keys is needed. (The research was funded by National Science Center, decision DEC-2012/07/D/ST6/02483.).