Year: 2025
Type: conference-paper
Abstract: This article investigates the application of NVIDIA GPU shared memory for enhancing query performance in Online Analytical Processing (OLAP) systems. It delves into the unique capabilities of GPU architectures, specifically focusing on how shared memory can be optimized to accelerate analytical database queries. The discussion encompasses the use of integer-based encoding for organizing dimension tables and aggregated structures as simple arrays, which aligns with the GPU’s memory banks to optimize data access and minimize overhead. Key strategies outlined include the dynamic resolution of memory bank conflicts through intelligent data reallocation and the management of multiple data copies within the GPU’s limited shared memory. This is particularly significant in handling data-intensive operations common in OLAP systems, where efficiency and speed are paramount. The article also explores methods to mitigate potential bank and warp conflicts by strategically placing and replicating data, thus ensuring optimal parallelism and throughput. Furthermore, we analyse the impact of data distribution patterns, including the handling of skewed data distributions, on memory access and query performance. By examining how data skew affects serialization and access efficiencies, the study provides insights into the optimal management of shared memory resources to enhance query response times and overall system performance. In conclusion, the article highlights the transformative potential of GPU shared memory in processing large-scale data queries within OLAP systems, proposing several techniques to exploit the parallel processing capabilities of GPUs effectively. These optimizations not only reduce query latencies but also scale up the analytical capabilities of databases to meet the challenges of modern data environments.
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4410317163