28078

Comparing SYCL data transfer strategies for tracking use cases

S.Joube, H. Grasland, D. Chamont, E.Brunet
Universite Paris-Saclay, CNRS/IN2P3, IJCLab, 91405 Orsay, France
Journal of Physics: Conference Series 2438, 012018, 2023

@inproceedings{joube2023comparing,

   title={Comparing SYCL data transfer strategies for tracking use cases},

   author={Joube, S and Grasland, H and Chamont, D and Brunet, E},

   booktitle={Journal of Physics: Conference Series},

   volume={2438},

   number={1},

   pages={012018},

   year={2023},

   organization={IOP Publishing}

}

Download Download (PDF)   View View   Source Source   

492

views

The aim of this work is to compare the performance and ease of programming of the various data transfer strategies provided by SYCL 2020: buffers/accessors on one hand and the different storage types exposed by Unified Shared Memory (USM) on the other hand. We measured the relative performance of USM exclusively located either on the host (USM host) or on the device (USM device), or automatically managed and moved (USM shared). We also tried to quantify the impact of formatting data in a GPU-friendly manner as opposed to using a regular pointer-based structure in which the C++ allocator was replaced by the SYCL allocator sycl::malloc. We first made a PCIe-bound micro-benchmark to test the SYCL memory models with a simple access pattern. We then switched to a real use-case provided by Traccc, a research project associated to the ACTS particle tracking library. The algorithm of interest is SparseCCL, a clustering algorithm used on the first step of track reconstruction. For consistency, all tests were made on a single hardware architecture: a recent server having a discrete GPU with dedicated VRAM, representative of the hardware currently used by the ACTS team.
No votes yet.
Please wait...

Recent source codes

* * *

* * *

HGPU group © 2010-2024 hgpu.org

All rights belong to the respective authors

Contact us: