Technical Papers Archive

From Error-Bounded to Fixed-Ratio: Efficient Sampling-Guided GPU Lossy Compression


Authors: Bohan Zhang (University of Florida), Yafan Huang (University of Iowa), Sheng Di (Argonne National Laboratory), and Guanpeng Li (University of Florida)

Abstract: Modern high performance computing (HPC) applications generate massive datasets, creating memory pressure and costly data movement that hinder scientific discovery and AI innovation. GPU lossy compression reduces data volume, but existing approaches face limitations. Error-bounded compression provides strict quality guarantees but produces unpredictable compressed sizes, complicating memory provisioning and system integration. Fixed-ratio compression ensures predictable memory requirements but lacks explicit error control, potentially compromising data fidelity. We introduce MILIO, a GPU lossy compression framework that combines fixed-ratio control with error-bounded compression. Built on dimension-aware sampling, fine-grained ratio profiling, and multi-stream global synchronization, MILIO achieves user-specified compression ratios with high throughput while supporting error control. Across nine real-world scientific datasets, MILIO achieves a compression throughput of 257.1 GB/s on an NVIDIA A100 GPU, delivering 3.25× the throughput of cuZFP at comparable compression ratios. It integrates seamlessly with scientific and AI workloads and demonstrates near-optimal scalability across multiple GPUs.


Presentation: file

Back to Technical Papers Archive Listing