TL;DR
This paper introduces MSz, a novel algorithm that corrects error-bounded lossy compressed data to preserve crucial topological features called Morse-Smale (MS) segmentations. By generating a set of "edits" during compression, MSz ensures that the decompressed data accurately reflects the original's topological structure, preventing misinterpretations in scientific analysis, and it achieves this efficiently through parallel CPU and GPU implementations.
What Problem Does This Paper Address?
Scientific simulations, especially in fields like fluid dynamics, climate science, and cosmology, generate massive datasets. To manage these, scientists rely on error-bounded lossy compressors to reduce data size while maintaining a certain level of accuracy. However, a significant problem is that these compressors often fail to preserve the topology of the data. Topological features, such as Morse-Smale (MS) segmentations, are critical for understanding scientific phenomena like chemical bonds, reaction zones, or atmospheric river structures. When these features are distorted by compression, it can lead to incorrect scientific conclusions. Existing lossy compressors don't typically account for these topological structures, creating a gap between data reduction efficiency and scientific interpretability.
Methods
MSz tackles this problem by introducing an "edit-based" paradigm. Instead of modifying the compression algorithm itself, MSz operates after the data has been compressed and decompressed. The core idea is to generate a set of small, targeted modifications (edits) to the decompressed data that, when applied, restore the original MS segmentation without violating the user-defined error bound.
The process involves two main iterative loops:
- Critical Point Loops (C-Loops): These loops identify and correct "false" critical points (minima and maxima) in the decompressed data. This involves iteratively adjusting scalar values to ensure that local extrema in the decompressed data match those in the original.
- Regular Point Loops (R-Loops): Once critical points are corrected, these loops focus on ensuring that the "integral lines" (paths tracing the flow of scalar values) correctly connect to their intended minima and maxima. This step corrects mislabeled regular points.
These C- and R-loops alternate until the MS segmentation is identical to the original. To make this computationally feasible, MSz employs efficient shared-memory (CPU) and GPU parallelism, addressing challenges like concurrent data modifications with atomic operations. The generated edits are then losslessly compressed and appended to the original compressed data.
Datasets
The paper evaluates MSz using a variety of datasets from different scientific domains:
- Climate: Red Sea dataset (3D flow field), Integrated Vapor Transport (IVT) dataset (atmospheric conditions).
- Cosmology: Nyx dataset (cosmological hydrodynamics simulation).
- Molecular Dynamics: Adenine Thymine dataset (electron density).
- Fluid Dynamics: Heated Flow dataset (2D flow around a cylinder), Viscous Fingering dataset (ensemble simulation of viscous fingers).
- Combustion: Combustion dataset (turbulent combustion processes).
- Earth Science: Earthquake dataset (seismic simulation).
These datasets represent diverse topological complexities and are used to demonstrate the effectiveness of MSz across different scientific applications.
Key Insights and Ideas
- Topology Preservation as a Post-Compression Step: The most significant insight is the development of a method that corrects existing lossy compression outputs for topology preservation, rather than requiring modification of the compressors themselves. This makes MSz broadly applicable to any error-bounded lossy compressor.
- Iterative Refinement for Topological Accuracy: The alternating C-loop and R-loop strategy is crucial. It highlights that correcting critical points can introduce errors in regular points, and vice-versa, necessitating an iterative approach to achieve convergence and guarantee topological correctness.
- Edits as a Compact Representation of Topological Correction: The generated edits are sparse and can be losslessly compressed, adding only a small overhead to the overall compressed file size while achieving significant topological fidelity.
- Parallelism is Key for Practicality: The paper demonstrates that achieving practical performance for this computationally intensive correction process relies heavily on efficient parallelization, particularly on GPUs.
Key Results
- Accurate MS Segmentation Preservation: MSz successfully preserves MS segmentations across various datasets and with different compressors (SZ3 and ZFP), even with small error bounds. Figure 1 shows that off-the-shelf compressors can distort MS segmentations by up to 100%, while MSz corrects these distortions.
- Significant Speedup with Parallelism: The parallel implementation achieves substantial acceleration. On an NVIDIA A100 GPU, MSz shows up to a 198x speedup for end-to-end performance compared to serial CPU execution, and up to 700x speedup for individual tasks.
- Reasonable Compression Ratios: While MSz adds edits, the overall compression ratios remain competitive, often outperforming lossless compression methods and offering a good balance between data reduction and feature preservation. For example, with the Nyx dataset and a 10^-6 error bound, MSz achieves a compression ratio of 7.76x, while GZIP is 1.17x.
- Error Bound vs. Edit Ratio: The size of the edits required is generally proportional to the global error bound. Larger error bounds in the original compression lead to more topological distortions and thus more edits.
Limitations and Future Work
- Computational Overhead: Preserving MS segmentations adds computational cost and storage overhead for the edits. Users need to balance data reduction, performance, and feature preservation needs.
- Saddle Point Preservation: The current method focuses on MS segmentations, which do not include saddle points. Preserving full MS complexes, including saddles, is a more complex problem that requires further research due to the challenges in guaranteeing convergence with iterative strategies.
- Generalization to Full MS Complex: The authors suggest extending the method to preserve the full MS complex, which would involve incorporating saddle point preservation.
- Edit Compression: While edits are losslessly compressed, further improvements in this compression could enhance the overall compression ratio.
- Hardware Utilization Evaluation: A more in-depth evaluation of hardware utilization, throughput, and power consumption for GPU implementations is suggested for future work.
Summary updated