• Error-bounded lossy compression tackles exploding scientific data volumes in HPC. • Pre‑quantization removes sequential dependency, enabling high parallelism but introduces artifacts. • Authors analyze correlation between quantization index and compression error. • Introduce quantization‑aware interpolation to reduce artifacts and improve data fidelity. • Algorithm parallelized for shared‑ and distributed‑memory systems, preserving throughput. • Experiments on two leading compressors and five real datasets show quality gains.
Article Summaries:
- Researchers have addressed quality issues in high‑throughput, error‑bounded scientific data compressors that use pre‑quantization. Pre‑quantization removes sequential dependencies to enable massive parallelism, but often introduces noticeable artifacts when user‑specified error bounds are moderate or large. The study first characterizes the relationship between quantization indices and compression errors, then introduces a quantization‑aware interpolation algorithm to reduce these artifacts. The method is implemented for both shared‑memory and distributed‑memory systems, preserving the compressors’ speed. Evaluation on two leading pre‑quantization compressors and five real‑world datasets shows significant quality gains with minimal impact on compression throughput.
Sources: