Skip to main navigation Skip to search Skip to main content

A Fast Parallel Selection Algorithm on GPUs

Darius Bakunas-Milanowski, Vernon Rego, Janche Sang, Chansu Yu

    Research output: Contribution to journalArticlepeer-review

    Abstract

    Today, parallel selection algorithms that run on Graphical Processing Units (GPUs) hold great promise in providing even more computational power than that of conventional CPUs. To quantify these gains, we examined a new parallel selection algorithm to see exactly what its vast number of simple, data parallel, multithreaded cores meant for performance times, using the current generation of NVIDIA GPUs. Specifically, our team tested how we could utilize a GPU to select elements from a massive array that met specific criteria and store their indices in a target array for additional processing. In this paper, we report optimization techniques and road blocks encountered. Overall, the experimental results demonstrate that our implementation performs an average of 3.67 times faster than Thrust, an open-source parallel algorithms library.

    Keywords

    • CUDA Thrust Library
    • GPU
    • Optimization Techniques
    • Parallel Selection
    • SIMT

    Disciplines

    • Computer Sciences

    Cite this