Authors: Darren Ng and Abdullah Naveed (University of Florida); Duo Zhang (University of California, Merced); Guanpeng Li (University of Florida); Sheng Di (Argonne National Laboratory); and Xiaoyi Lu (University of Florida)
Abstract: Lossy compression is widely used to reduce storage space and I/O demands. However, its benefits diminish on NVMe-SSDs, where compression and decompression runtimes can exceed raw device throughput. Leveraging the low-level Storage Performance Development Kit (SPDK) driver to bypass the kernel and filesystem overheads, we conduct the first systematic study of lossy compressed I/O on NVMe-SSDs and expose key bottlenecks in process control, pipeline utilization, and parameter tuning. To address these issues, we design Z-SPDK, a novel extension of SPDK that incorporates fixed-ratio compression control, dual-pipeline parallelism, per-chunk error-bound, and zero-copy data transfer, thereby co-optimizing lossy compression with NVMe I/O. Experiments on three high-speed lossy compressors across five scientific datasets demonstrate up to 12X runtime speedup, while satisfying user-defined error bounds and improving NVMe bandwidth utilization by up to 5.9X. In addition, we evaluate with HDF5 and observe up to 82.56% runtime reduction with Z-SPDK.