Exploiting Graphics Processing Units for Massively Parallel Multi-Dimensional Indexing

Abstract

Scientific applications process truly large amounts of multi-dimensional datasets. To efficiently navigate such datasets, various multi-dimensional indexing structures, such as the R-tree, have been extensively studied for the past couple of decades. Since the GPU has emerged as a new cost-effective performance accelerator, now it is common to leverage the massive parallelism of the GPU in various applications such as medical image processing, computational chemistry, and particle physics. However, hierarchical multi-dimensional indexing structures are inherently not well suited for parallel processing because their irregular memory access patterns make it difficult to exploit massive parallelism. Moreover, recursive tree traversal often fails due to the small run-time stack and cache memory in the GPU. First, we propose Massively Parallel Three-phase Scanning (MPTS) R-tree traversal algorithm to avoid the irregular memory access patterns and recursive tree traversal so that the GPU can access tree nodes in a sequential manner. The experimental study shows that MPTS R-tree traversal algorithm consistently outperforms traditional recursive R-Tree search algorithm for multi-dimensional range query processing. Next, we focus on reducing the query response time and extending n-ary multi-dimensional indexing structures - R-tree, so that a large number of GPU threads cooperate to process a single query in parallel. Because the number of submitted concurrent queries in scientific data analysis applications is relatively smaller than that of enterprise database systems and ray tracing in computer graphics. Hence, we propose a novel variant of R-trees Massively Parallel Hilbert R-Tree (MPHR-Tree), which is designed for a novel parallel tree traversal algorithm Massively Parallel Restart Scanning (MPRS). The MPRS algorithm traverses the MPHR-Tree in mostly contiguous memory access patterns without recursion, which offers more chances to optimize the parallel SIMD algorithm. Our extensive experimental results show that the MPRS algorithm outperforms the other stackless tree traversal algorithms, which are designed for efficient ray tracing in computer graphics community. Furthermore, we develop query co-processing scheme that makes use of both the CPU and GPU. In this approach, we store the internal and leaf nodes of upper tree in CPU host memory and GPU device memory, respectively. We let the CPU traverse internal nodes because the conditional branches in hierarchical tree structures often cause a serious warp divergence problem in the GPU. For leaf nodes, the GPU scans a large number of leaf nodes in parallel based on the selection ratio of a given range query. It is well known that the GPU is superior to the CPU for parallel scanning. The experimental results show that our proposed multi-dimensional range query co-processing scheme improves the query response time by up to 12x and query throughput by up to 4x compared to the state-of-the-art GPU tree traversal algorithm.