High-Throughput Data Structures for GPU-Accelerated Computing
Karthik Reddy · International Journal of Artificial Intelligence Data Science and Machine Learning · 2022
The increasing demand for high-performance computing in fields like scientific simulations, machine learning, and data analysis has driven the adoption of Graphics Processing Units (GPUs) as accelerators. GPUs offer massive parallelism, but effectively harnessing their power requires careful consideration of data structures. Traditional CPUcentric data structures often become bottlenecks when deployed in GPU environments due to memory access patterns and synchronization overhead. This paper explores the landscape of high-throughput data structures specifically designed for GPU-accelerated computing. We discuss key considerations for GPU data structure design, including memory layout, access patterns, concurrency management, and data transfer strategies. We then delve into specific data structures optimized for GPU execution, such as array of structures vs. structure of arrays, sparse matrix formats, tree-based structures (e.g., B-trees), and hash tables. We analyze their performance characteristics, trade-offs, and suitability for different application domains. Finally, we present case studies demonstrating the effectiveness of these data structures in real-world GPU-accelerated applications and discuss future research directions in this critical area