[Paper Review] SZx: an Ultra-fast Error-bounded Lossy Compressor for Scientific Datasets

Xiaodong Yu, Sheng Di+5 more|arXiv (Cornell University)|Jan 31, 2022

Advanced Data Storage Technologies6 citations

TL;DR

This paper proposes UFZ, an ultra-fast error-bounded lossy compressor for scientific datasets that uses only lightweight operations like bitwise, addition, and subtraction to achieve extreme compression and decompression speeds on both CPU and GPU. UFZ achieves up to 264GB/s compression and 446GB/s decompression on GPU, outperforming SZ and ZFP by 2–16× in speed while maintaining acceptable compression ratios and strict error bounds.

ABSTRACT

Today's scientific high performance computing (HPC) applications or advanced instruments are producing vast volumes of data across a wide range of domains, which introduces a serious burden on data transfer and storage. Error-bounded lossy compression has been developed and widely used in scientific community, because not only can it significantly reduce the data volumes but it can also strictly control the data distortion based on the use-specified error bound. Existing lossy compressors, however, cannot offer ultra-fast compression speed, which is highly demanded by quite a few applications or use-cases (such as in-memory compression and online instrument data compression). In this paper, we propose a novel ultra-fast error-bounded lossy compressor, which can obtain fairly high compression performance on both CPU and GPU, also with reasonably high compression ratios. The key contributions are three-fold: (1) We propose a novel, generic ultra-fast error-bounded lossy compression framework -- called UFZ, by confining our design to be composed of only super-lightweight operations such as bitwise and addition/subtraction operation, still keeping a certain high compression ratio. (2) We implement UFZ on both CPU and GPU and optimize the performance according to their architectures carefully. (3) We perform a comprehensive evaluation with 6 real-world production-level scientific datasets on both CPU and GPU. Experiments show that UFZ is 2~16X as fast as the second-fastest existing error-bounded lossy compressor (either SZ or ZFP) on CPU and GPU, with respect to both compression and decompression.

Motivation & Objective

• Address the critical need for ultra-high-speed lossy compression in scientific HPC workloads, especially for in-memory and real-time instrument data processing.
• Overcome the performance bottleneck of existing error-bounded compressors like SZ and ZFP, which rely on expensive operations (e.g., multiplication, division) that limit throughput.
• Design a compression framework that uses only super-lightweight operations (bitwise, add, subtract) while still maintaining strict error bounds and reasonable compression ratios.
• Optimize the UFZ framework for both CPU and GPU architectures through device-specific performance tuning and memory access patterns.
• Demonstrate the practical superiority of UFZ in real-world HPC environments using production-level scientific datasets on supercomputers like Summit and ThetaGPU.

Proposed method

• Design a novel compression framework, UFZ, that relies exclusively on lightweight operations—bitwise shifts, additions, and subtractions—avoiding costly multiplications and divisions.
• Implement UFZ on both CPU and GPU, leveraging SIMD parallelism on CPU and thread-level parallelism on GPU for high throughput.
• Integrate a lightweight error-control mechanism that ensures prediction errors remain within user-specified bounds using integer arithmetic and fixed-point scaling.
• Use a simplified quantization scheme based on error thresholding and rounding, avoiding complex operations like division in per-point quantization.
• Optimize memory access patterns and kernel launches for GPU, minimizing latency and maximizing occupancy and memory bandwidth utilization.
• Apply a hierarchical compression strategy that processes data in small, cache-friendly blocks to improve data locality and parallel efficiency.

Experimental results

Research questions

• RQ1 Can a lossy compression framework achieve ultra-high throughput on CPU and GPU by relying solely on lightweight operations such as bitwise and arithmetic operations?
• RQ2 How does the performance of UFZ compare to state-of-the-art compressors (SZ and ZFP) in terms of compression and decompression speed across different hardware platforms?
• RQ3 To what extent does UFZ maintain acceptable compression ratios while ensuring strict error bounds in real-world scientific datasets?
• RQ4 How does UFZ impact end-to-end I/O performance when used in data dumping and loading pipelines on large-scale HPC systems?
• RQ5 Can UFZ be efficiently optimized for heterogeneous architectures like CPU and GPU without sacrificing error control or compression quality?

Key findings

• UFZ achieves a peak compression throughput of 264GB/s on a single GPU (A100), which is 2–16× faster than the second-fastest compressor (SZ or ZFP) on both CPU and GPU.
• On GPU, UFZ reaches a peak decompression speed of 446GB/s, significantly outperforming cuSZ and cuZFP, which achieve only 9.7–67GB/s.
• On CPU, UFZ is 2.5–5× faster than ZFP and 5–7× faster than SZ in compression, and 2–4× faster in decompression.
• In end-to-end data I/O pipelines on ANL ThetaGPU with 64–1024 cores, UFZ improves data dumping and loading performance by 100–200% compared to SZ and ZFP.
• UFZ achieves reasonable compression ratios (0.3–3× lower than ZFP and 3–30× lower than SZ) despite avoiding complex prediction and transform steps.
• The performance gains are attributed to the elimination of expensive operations and the use of highly optimized, architecture-aware kernels on both CPU and GPU.