L3: A GPU-Native Co-Designed Data Format for Learned Lossless Lightweight Compression
Open Access
- 18 May 2026
- journal article
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (3) , 1-27
Abstract
Learned Compression achieves strong CPU performance but lacks a GPU-native format, limiting its use in GPU analytics. We present L3, a GPU-native Learned Lossless Lightweight Compression format that enables end-to-end on-device processing with efficient compression, high-throughput decompression, and fast random access on GPU. On NVIDIA GPUs, a warp is a group of 32 threads; we refer to each thread as a lane (lane id 0–31), and call a layout lane-major when each lane's words are stored contiguously. L3 introduces three tightly coupled components built around the SLAP Vertical layout. First, the L3 Storage Layout (SLAP) stores bit-packed residual streams in a lane-major organization, i.e., residual words are laid out lane by lane so each warp lane consumes a contiguous word sequence in memory, exploiting the GPU L1 sector cache for implicit prefetching and high reuse during unpacking. Second, the Warp-Cooperative Learned Decompression Module maps each partition to one thread block and decodes warp tiles using per-lane bit readers, branchless bit extraction, and a bit-exact no-FMA FP64 finite-difference predictor. Third, the GPU-Native Learned Compression Pipeline builds adaptive partitions via bulk delta-bits analysis, scan/compaction, and an odd-even GPU merge loop, then packs residuals directly into the final SLAP Vertical layout on the device. L3 achieves high performance on modern GPUs. It encodes 3–6× faster than Tile and FastLanes-GPU and sustains 1.08–1.90 TB/s decompression throughput, comparable to the fastest lightweight GPU codecs. On correlated datasets, L3 reaches up to 77× compression while remaining competitive on weakly correlated inputs. For random access, L3 maintains 1.2–2.6 Billion queries/s and outperforms Tile-DFOR/Tile-RFOR by 5–10×. On SSB with unified query plans, L3 achieves the lowest average latency (1.14 ms), matching or outperforming state-of-the-art GPU baselines.Keywords
Funding Information
- Beijing Science and Technology Project (Z251100008125032)
This publication has 56 references indexed in Scilit:
- Learning database optimization techniques: the state-of-the-art and prospectsFrontiers of Computer Science, 2025
- GaussDB-AISQL: a composable cloud-native SQL system with AI capabilitiesFrontiers of Computer Science, 2025
- L2SM: a query-optimized linked LSM-tree for HTAP workloadsFrontiers of Computer Science, 2024
- TADOC: Text analytics directly on compressionThe VLDB Journal, 2020
- Decoding billions of integers per second through vectorizationSoftware: Practice and Experience, 2013
- A Difference Fitting Residuals algorithm for lossless data compression in wireless sensor nodes2011
- Speeding Up Queries in Column StoresLecture Notes in Computer Science, 2010
- The Star Schema Benchmark and Augmented Fact Table Indexing2009
- Program design for retrospective searches on large data basesInformation Storage and Retrieval, 1972
- A Mathematical Theory of CommunicationBell System Technical Journal, 1948