Scholarlink

<p>Adaptive mesh refinement (AMR) has become an indispensable tool in high-performance computing (HPC), enabling exascale simulations by dynamically allocating computational resources and significantly reducing memory footprint. Meanwhile, lossy compression is widely adopted in HPC environments to alleviate critical storage capacity and I/O bottlenecks, provided that reconstruction errors remain within acceptable bounds. However, the hierarchical structure, multi-resolution nature, and inherent spatio-temporal irregularity of AMR data pose unique challenges that render general-purpose compressors inefficient. Despite their respective advantages, existing data compressors still have an insufficient compression ratio and low throughput for data reduction in AMR applications. This paper mainly explores how to improve the performance of state-of-the-art lossy compression algorithms from the perspective of applications. To this end, we propose a density-partitioned AMR data lossy compression framework called AMRDPC, improving AMR applications’ storage efficiency. The main ideas are twofold. First, to address the high computational overhead of using the k-d tree to process medium-density AMR data, we propose a fast k-d tree backfilling density grid (FBKDTree) strategy to improve compression speed. Second, to address the problem of the low compression ratio of high-density AMR data, we propose an efficient loop reversal patching (ELRP) strategy based on the design characteristics of existing prediction-based compressors. It can significantly improve the data compression performance while controlling errors. To verify the effectiveness of AMRDPC, we introduce multiple evaluation metrics for experimental analysis in seven real AMReX application datasets. Compared to state-of-the-art methods, AMRDPC achieves significant performance gains, with up to a 5.73<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> higher compression ratio and an 18.83% increase in throughput, providing a powerful data reduction solution for supercomputing environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An efficient lossy compression framework for density partitioning in AMR applications

  • 期刊论文

  • 发表日期：

    2025年12月6日

查看全文
  • Yida Li,
  • Huizhang Luo,
  • Yufeng Zhang,
  • Keqin Li,
  • Kenli Li

摘要

Adaptive mesh refinement (AMR) has become an indispensable tool in high-performance computing (HPC), enabling exascale simulations by dynamically allocating computational resources and significantly reducing memory footprint. Meanwhile, lossy compression is widely adopted in HPC environments to alleviate critical storage capacity and I/O bottlenecks, provided that reconstruction errors remain within acceptable bounds. However, the hierarchical structure, multi-resolution nature, and inherent spatio-temporal irregularity of AMR data pose unique challenges that render general-purpose compressors inefficient. Despite their respective advantages, existing data compressors still have an insufficient compression ratio and low throughput for data reduction in AMR applications. This paper mainly explores how to improve the performance of state-of-the-art lossy compression algorithms from the perspective of applications. To this end, we propose a density-partitioned AMR data lossy compression framework called AMRDPC, improving AMR applications’ storage efficiency. The main ideas are twofold. First, to address the high computational overhead of using the k-d tree to process medium-density AMR data, we propose a fast k-d tree backfilling density grid (FBKDTree) strategy to improve compression speed. Second, to address the problem of the low compression ratio of high-density AMR data, we propose an efficient loop reversal patching (ELRP) strategy based on the design characteristics of existing prediction-based compressors. It can significantly improve the data compression performance while controlling errors. To verify the effectiveness of AMRDPC, we introduce multiple evaluation metrics for experimental analysis in seven real AMReX application datasets. Compared to state-of-the-art methods, AMRDPC achieves significant performance gains, with up to a 5.73 \(\times \) × higher compression ratio and an 18.83% increase in throughput, providing a powerful data reduction solution for supercomputing environments.