Co-processing heterogeneous parallel index for multi-dimensional datasets

Abstract

We present a novel multi-dimensional range query co-processing scheme for the CPU and GPU. It has been reported that traversing hierarchical tree structures in parallel is inherently not efficient because of large branching factors. Besides, it is known that the recursive tree traversal algorithm required for multi-dimensional range queries is not well suited for the GPU architecture owing to its small shared memory.

In this paper, we propose co-processing range queries using both the CPU and GPU to make the most use of each architecture. In Hybrid tree that we present in this paper, we let CPU navigate the internal nodes of hierarchical tree structures and make GPU scan leaf nodes in a linear fashion using a massively large number of processing units. With the co-processing scheme, we can asynchronously leverage the strengths of each architecture. We also propose a novel dynamic GPU block scheduling algorithm for multiple range queries. In our scheduling algorithm, we consider the selection ratio of each query to determine the number of GPU blocks to launch. By assigning the right number of GPU blocks, we can significantly improve the query processing throughput for multiple concurrent queries. Our extensive experimental study shows that the proposed co-processing scheme shows up to 12× faster query response time than the state-of-the-art GPU tree traversal algorithm. We also show that our dynamic GPU block assignment algorithm improves the query processing throughput by up to 4× .