Heterogeneous computing for accelerating structure query processing

Yong, Keh Kok (2026) Heterogeneous computing for accelerating structure query processing. Master dissertation/thesis, UTAR.

PDF Download (2942Kb)

Abstract

Analytical structured query language (SQL) workloads increasingly run on commodity servers, yet it remains unclear when a single graphics processing unit (GPU) can replace or augment small clusters. Existing GPU based database systems often optimise isolated kernels, assume uncompressed data, or offload only selected operators, so they give limited insight into end-to-end behaviour and the cost of moving data between CPU and GPU. This thesis establishes when a single GPU can accelerate SQL analytics. It develops a GPU structured query accelerator with a GPU-accelerated query pipeline that compresses, stores, schedules and processes data on the device as a first-class execution target. The design combines a compression-first preprocessing path based on a GPU PFOR-Delta variant, chunked columnar storage sized to device memory, an SQL front-end with a placement scheduler, and GPU centric hash and sort-merge joins that run directly on compressed chunks. It also integrates GPU implementations of exact string matching, histogram-based string predicates and short-pattern approximate matching so that structured and text conditions execute within one framework. The thesis implements this accelerator and evaluates it on synthetic joins of up to 512 million tuples, on TPC-H queries of up to 50 GB, and on healthcare and geospatial heat map workloads. The experiments use PostgreSQL on a multi-core CPU and an eight-node Hadoop/Spark cluster as baselines. PFORDelta reduces analytical data to 42–83% of its original size while it remains fast enough for pre-processing, and GPU joins reduce join times relative to the CPU at all tested scales. On TPC-H Query 2 at 50 GB, the single-GPU workstation achieves end-to-end speedups of 21–25x over PostgreSQL after PCI Express (PCIe) transfer costs, and on the healthcare and geospatial workloads it matches or exceeds the eight-node cluster, including 4–18x speedups for heat map generation. The evaluation maps data sizes and workload shapes that benefit most from GPU acceleration and identifies cases where PCIe transfers and memory bandwidth limits erode these gains. The evaluation maps data sizes and workload shapes that benefit most from GPU acceleration and identifies cases where PCIe transfers and memory bandwidth limits erode these gains. Together, the architecture and results establish architecture aware criteria for treating a single GPU as the primary analytical engine and provide a blueprint for GPU accelerated SQL systems that balance single GPU accelerators against multi core CPUs and small clusters.

Actions (login required)