[Paper Review] Understanding GPU-Based Lossy Compression for Extreme-Scale Cosmological Simulations

Sian Jin, Pascal Grosset+5 more|arXiv (Cornell University)|Apr 1, 2020

Computational Physics and Python Applications4 citations

TL;DR

This paper proposes GPU-based lossy compression for extreme-scale cosmological simulations using the Foresight framework to evaluate compressors like cuZFP and cuSZ. It achieves 5–15× compression ratios with high throughput and preserves post-analysis fidelity, especially for cosmology-specific metrics like power spectrum and halo finder results.

ABSTRACT

To help understand our universe better, researchers and scientists currently run extreme-scale cosmology simulations on leadership supercomputers. However, such simulations can generate large amounts of scientific data, which often result in expensive costs in data associated with data movement and storage. Lossy compression techniques have become attractive because they significantly reduce data size and can maintain high data fidelity for post-analysis. In this paper, we propose to use GPU-based lossy compression for extreme-scale cosmological simulations. Our contributions are threefold: (1) we implement multiple GPU-based lossy compressors to our opensource compression benchmark and analysis framework named Foresight; (2) we use Foresight to comprehensively evaluate the practicality of using GPU-based lossy compression on two real-world extreme-scale cosmology simulations, namely HACC and Nyx, based on a series of assessment metrics; and (3) we develop a general optimization guideline on how to determine the best-fit configurations for different lossy compressors and cosmological simulations. Experiments show that GPU-based lossy compression can provide necessary accuracy on post-analysis for cosmological simulations and high compression ratio of 5~15x on the tested datasets, as well as much higher compression and decompression throughput than CPU-based compressors.

Motivation & Objective

• Address the growing I/O and storage bottlenecks in extreme-scale cosmological simulations that generate petabytes of data.
• Overcome the limitations of CPU-based lossy compression, which introduces significant performance overhead (e.g., >10% of total time) on modern supercomputers.
• Develop a comprehensive evaluation framework that includes both general compression metrics and cosmology-specific analysis to ensure data fidelity for scientific post-processing.
• Provide a systematic optimization guideline to select the best-fit compression configuration for GPU-based lossy compressors in cosmological simulations.
• Enable in-situ compression with minimal performance impact while maintaining scientific accuracy for key astrophysical features like dark matter halos and power spectra.

Proposed method

• Implement GPU-accelerated versions of lossy compressors (cuZFP and cuSZ) within the open-source Foresight framework for benchmarking and analysis.
• Use Foresight’s PAT (Post-Analysis Toolkit) to evaluate data fidelity using cosmology-specific metrics, including power spectrum analysis and halo finder results.
• Measure compression and decompression throughput, including GPU-CPU data transfer overhead, to assess performance scalability on leadership-class systems.
• Systematically vary compression parameters (e.g., error bounds, block sizes) to explore trade-offs between compression ratio, distortion, and throughput.
• Integrate both general metrics (e.g., PSNR, MSE) and domain-specific metrics to validate that reconstructed data remain scientifically usable.
• Develop a three-step optimization guideline: (1) benchmark compressors across configurations, (2) validate fidelity using cosmology-relevant analysis, and (3) select the configuration with the highest compression ratio that maintains acceptable distortion.

Experimental results

Research questions

• RQ1 Can GPU-based lossy compression achieve high compression ratios (e.g., 5–15×) while maintaining sufficient data fidelity for cosmological post-analysis?
• RQ2 How does GPU-based compression compare to CPU-based compression in terms of throughput and I/O performance overhead in extreme-scale simulations?
• RQ3 Which compression configurations preserve key cosmological features such as the power spectrum and halo properties after reconstruction?
• RQ4 What is the optimal configuration of error bounds and block sizes that maximizes compression ratio without degrading scientific accuracy?
• RQ5 Can a unified optimization guideline be established to help cosmology simulation users select the best-fit GPU-based lossy compressor for their specific simulation data?

Key findings

• GPU-based lossy compression achieves compression ratios of 5–15× on real HACC and Nyx simulation datasets while maintaining high fidelity for cosmological analysis.
• cuZFP and cuSZ demonstrate significantly higher compression and decompression throughput than CPU-based versions, reducing I/O overhead and avoiding >10% performance degradation seen in prior CPU-based approaches.
• The power spectrum of reconstructed data remains within acceptable error bounds across all tested configurations, indicating preservation of large-scale structure information.
• Halo finder results show minimal deviation between original and reconstructed datasets, confirming that dark matter halo properties are preserved under optimal compression settings.
• The proposed optimization guideline successfully identifies the best-fit configuration by selecting the highest compression ratio setting that maintains acceptable distortion in both power spectrum and halo finder metrics.
• Throughput measurements show that GPU-based compression kernels achieve up to ~2 TB/s on Summit, with end-to-end throughput (including CPU-GPU transfer) still exceeding 1 TB/s, making in-situ compression viable for extreme-scale simulations.