EXaCTz: Guaranteed Extremum Graph and Contour Tree Preservation for Distributed- and GPU-Parallel Lossy Compression

EXaCTz is a parallel algorithm designed for distributed and GPU-parallel lossy compression that guarantees the preservation of both extremum graphs and contour trees in scalar field data. It achieves this by enforcing localized scalar ordering constraints rather than relying on explicit topology construction, which allows it to bridge the significant performance gap between topology correction speeds (typically MB/s) and compression speeds (GB/s).

Key performance metrics for EXaCTz include: Single-GPU Throughput: Up to 4.52 GB/s, which is up to 213x faster than state-of-the-art CPU-based methods and 3,285x faster than previous GPU implementations. Distributed Scaling: Scales to 128 GPUs with 55.6% efficiency, processing 512 GB datasets in under 48 seconds with an aggregate throughput of 32.69 GB/s. * Accuracy: Achieves 100% preservation of extremum graphs and contour trees across tested datasets, outperforming methods like TopoA which may leave residual distortions.

The algorithm was authored by Yuxiao Li, Mingze Xia, Xin Liang, Bei Wang, and Hanqi Guo from The Ohio State University and Oregon State University, with the preprint submitted to arXiv in April 2026.

EXaCTz is a parallel algorithm for correcting lossy-compressed scalar field data so that important topological structures are retained after compression. It targets a common problem in large-scale scientific data management: error-bounded compression can reduce storage and I/O costs, but even small numerical perturbations may destroy critical features such as extrema, saddle connections, and contour-tree relationships. Rather than preserving only one topological representation, EXaCTz aims to enforce consistency for both extremum graphs and contour trees at the same time, making the compressed data more faithful for topology-sensitive analysis.

A central contribution of the paper is the design of a scalable correction method that can run efficiently on GPUs and distributed-memory systems. Existing topology-preserving compression approaches often face performance bottlenecks, limited parallelism, or trade-offs between compression throughput and topological fidelity. EXaCTz addresses this by formulating the correction problem in a way that can be parallelized across modern high-performance computing architectures, enabling topology preservation to remain practical even for very large simulation datasets.

The work matters because scalar-field topology is a primary basis for feature extraction, visualization, and downstream scientific interpretation in areas such as fluid dynamics, climate modeling, astrophysics, and computational mechanics. By making simultaneous extremum-graph and contour-tree preservation feasible in parallel, EXaCTz helps bridge the gap between aggressive lossy compression and scientifically meaningful data fidelity. In practice, this could allow researchers to compress and store massive simulation outputs more efficiently without silently corrupting the topological features that make those outputs analytically useful.