Logo image
Back
cuSZ-I: High-Fidelity Error-Bounded Lossy Compression for Scientific Data on GPUs
Preprint   Open access

cuSZ-I: High-Fidelity Error-Bounded Lossy Compression for Scientific Data on GPUs

Jinyang Liu, Jiannan Tian, Shixun Wu, Sheng Di, Boyuan Zhang, Yafan Huang, Kai Zhao, Guanpeng Li, Dingwen Tao, Zizhong Chen, …
arXiv.org
Cornell University
12/09/2023
DOI: 10.48550/arxiv.2312.05492
url
https://doi.org/10.48550/arxiv.2312.05492View
Preprint (Author's original)This preprint has not been evaluated by subject experts through peer review. Preprints may undergo extensive changes and/or become peer-reviewed journal articles. Open Access

Abstract

Error-bounded lossy compression is a critical technique for significantly reducing scientific data volumes. Compared to CPU-based scientific compressors, GPU-accelerated compressors exhibit substantially higher throughputs, which can thus better adapt to GPU-based scientific simulation applications. However, a critical limitation still lies in all existing GPU-accelerated error-bounded lossy compressors: they suffer from low compression ratios, which strictly restricts their scope of usage. To address this limitation, in this paper, we propose a new design of GPU-accelerated scientific error-bounded lossy compressor, namely cuSZ-I, which has achieved the following contributions: (1) A brand new GPU-customized interpolation-based data pre-diction method is raised in cuSZ-I for extensively improving the compression ratio and the decompression data quality. (2) The Huffman encoding module in cuSZ-I has been improved for both efficiency and stability. (3) cuSZ-I is the first work to integrate the highly effective NVIDIA bitcomp lossless compression module to maximally boost the compression ratio for GPU-accelerated lossy compressors with nearly negligible speed degradation. In experimental evaluations, with the same magnitude of compression throughput as existing GPU-accelerated compressors, in terms of compression ratio and quality, cuSZ-I outperforms other state-of-the-art GPU-based scientific lossy compressors to a significant extent. It gains compression ratio improvements by up to 500% under the same error bound or PSNR. In several real-world use cases, cuSZ-I also achieves the optimized performance, having the minimized time cost for distributed lossy data transmission tasks and the highest decompression data visualization quality.
 Computer Science - Distributed, Parallel, and Cluster Computing

Details

Metrics

10 Record Views
Logo image