Are van Emde Boas trees viable on the GPU?

Van Emde Boas trees show an asymptotic query complexity surpassing the performance of traditional data structure for performing search queries in large data sets. However, their implementation and large constant overheads prohibit their widespread use. In this work, we ask, whether van Emde Boas trees are viable on the GPU. We presents a novel algorithm to construct a van Emde Boas tree utilizing the parallel compute power and memory bandwidth of modern GPUs. By analyzing the structure of a sorted data set, our method is able to build a van Emde Boas tree efficiently in parallel with little thread synchronization. We compare querying data using a van Emde Boas tree and binary search on the GPU and show that for very large data sets, the van Emde Boas tree outperforms a binary search by up to 1.2x while similarly increasing space requirements. Overall, we confirm that van Emde Boas trees are viable for certain scenarios on the GPU.