FAST '26
as the organisers put it
24th USENIX Conference on File and Storage Technologies (FAST '26), February 24–26, 2026, Hyatt Regency Santa Clara, Santa Clara, CA, USA. FAST brings together storage-system researchers and practitioners to explore new directions in the design, implementation, evaluation, and deployment of storage systems.
| title | author | length |
|---|---|---|
| Here, There and Everywhere: The Past, the Present and the Future of Local Storage in... | Leping Yang, Yanbo Zhou, Gong Zeng, Li Zhang, Saisai Zhang, Ruilin Wu, Chaoyang Sun, Shiyi Luo, Wenrui Li, Keqiang Niu, Xiaolu Zhang, Junping Wu, Jiaji Zhu, Jiesheng Wu, Mariusz Barczak, Wayne Gao, Ruiming Lu, Erci Xu, Guangtao Xue | 15m |
Cloud local storage evolution is traced from Espresso (SPDK user-space poll), which cuts kernel context-switch overhead but wastes CPU (99% of nodes show <60%… | ||
| Cost-efficient Archive Cloud Storage with Tape: Design and Deployment | Qing Wang, Fan Yang, Qiang Liu, Geng Xiao, Yongpeng Chen, Hao Lan, Leiming Chen, Bangzhu Chen, Chenrui Liu, Pingchang Bai, Bin Huang, Zigan Luo, Mingyu Xie, Yu Wang, Youyou Lu, Huatao Wu, Jiwu Shu | 14m |
TapeOBS claims tape reduces TCO 4.95× and CO2e by 97% and runs in production storing hundreds of petabytes since 2024. It uses a 4% HDD staging pool to make a… | ||
| SolidAttention: Low-Latency SSD-based Serving on Memory-Constrained PCs | Xinrui Zheng, Dongliang Wei, Jianxiang Gao, Yixin Song, Zeyu Mi, Haibo Chen | 19m |
SolidAttention is an LLM inference engine that offloads the full KV cache to SSD by consolidating KV pairs into coarse-grained blocks and using speculative… | ||
| CacheSlide: Unlocking Cross Position-Aware KV Cache Reuse for Accelerating LLM Serving | Yang Liu, Yunfei Gu, Liqiang Zhang, Chentao Wu, Guangtao Xue, Jie Li, Minyi Guo, Junhao Hu, Jie Meng | 17m |
CacheSlide: Unlocking Cross Position-Aware KV Cache Reuse for Accelerating LLM Serving
Yang Liu and Yunfei Gu, Shanghai Jiao Tong University; Liqiang Zhang… | ||
| Bidaw: Enhancing Key-Value Caching for Interactive LLM Serving via Bidirectional... | Shipeng Hu, Guangyan Zhang, Yuqi Zhou, Yaya Wei, Ziyan Zhong, Jike Chen | 15m |
Bidaw implements a two-tier KV cache (host memory + SSD) and coordinates compute and storage: IO-aware request scheduling separates and reorders requests by… | ||
| Accelerating Model Loading in LLM Inference by Programmable Page Cache | Yubo Liu, Hongbo Li, Xiaojia Huang, Yongfeng Wang, Hanjun Guo, Hui Chen, Yuxin Ren, Ning Jia | 20m |
PPC implements a user-space, file-routing page cache that intercepts read events and runs per-model policies for prefetch and eviction; MAIO uses an… | ||
| OdinANN: Direct Insert for Consistently Stable Performance in Billion-Scale Graph-Based.. | Hao Guo, Youyou Lu | 12m |
OdinANN implements direct insert into an on-disk graph index (2x SSD overprovisioning) and claims billion-scale stability while ingesting up to 100 million… | ||
| DMTree: Towards Efficient Tree Indexing on Disaggregated Memory via Compute-side... | Guoli Wei, Yongkun Li, Haoze Song, Tao Li, Lulu Yao, Yinlong Xu, Heming Cui | 13m |
Prototype DMTree implements a compute-side collaborative B‑tree: leaf nodes keep key/value entries and per‑leaf fingerprint tables whose addresses live in… | ||
| An Efficient Cloud Storage Model with Compacted Metadata Management for Performance... | Kai Zhang, Tianyu Wang, Zili Shao | 15m |
Implements a prototype called CloudTS that separates metadata from compressed time-series objects on Amazon S3, builds a global tag dictionary (Patricia-trie… | ||
| "Range as a Key" is the Key! Fast and Compact Cloud Block Store Index with RASK | Haoru Zhao, Mingkai Dong, Erci Xu, Zhongyu Wang, Haibo Chen | 15m |
RASK is a tree-structured cloud block-store index that natively uses range-as-a-key with adaptive radix tree internals and B-tree cell leaves, a log-structured… | ||
| Holistic and Automated Task Scheduling for Distributed LSM-tree-based Storage | Yuanming Ren, Siyuan Sheng, Zhang Cao, Yongkun Li, Patrick P. C. Lee | 16m |
HATS is a scheduler implemented atop Cassandra that co-schedules reads and compactions, using coarse-grained replica selection (epochal 60s global balancing… | ||
| Keynote Address: Everything You Always Wanted to Know about Storage Analysis... | Erez Zadok | 49m |
Thirty years of storage work distills into concrete advice: expect thousands of experiments (a small matrix of micro/macro/trace runs plus varied sizes and… | ||
| Preparation Meets Opportunity: Enhancing Data Preprocessing for ML Training With Seneca | Omkar Desai, Ziyang Jiao, Shuyi Pei, Janki Bhimani, Bryan S. Kim | 18m |
Seneca is a modified PyTorch dataloader that uses model-driven partitioning to split DRAM between encoded, decoded, and augmented forms and opportunistic data… | ||
| Fast Cloud Storage for AI Jobs via Grouped I/O API with Transparent Read/Write... | Yingyi Hao, Ting Yao, Xingda Wei, Dingyan Zhang, Tianle Sun, Yiwen Zhang, Zhiyong Fu, Huatao Wu, Rong Chen | 17m |
AITURBO is a cloud storage system that cuts AI checkpoint IO by using the grouped I/O API to collect per-tensor metadata, derive an optimized write plan… | ||
| AdaCheck: An Adaptive Checkpointing System for Efficient LLM Training with Redundancy... | Weijie Liu, Shengwei Li, Zhiquan Lai, Keshi Ge, Qiaoling Chen, Peng Sun, Dongsheng Li, Kai Lu | 15m |
AdaCheck is an adaptive checkpointing system that minimizes checkpoint size by identifying tensor redundancy across parallelisms, model architectures, and… | ||
| Sharpen the Spec, Cut the Code: A Case for Generative File System with SYSSPEC | Qingyuan Liu, Mo Zou, Hengbin Zhang, Dong Du, Yubin Xia, Haibo Chen | 16m |
Claims a working pipeline that uses SYSSPEC to generate a concurrent file system called SPECFS from high-level formal-style specs, producing ~4,000 lines of C… | ||
| Cylon: Fast and Accurate Full-System Emulation of CXL-SSDs | Dongha Yoon, Hansen Idden, Jinshu Liu, Berkay Inceisci, Sam H. Noh, Huaicheng Li | 17m |
Cylon is a full-system CXL-SSD emulator that reproduces sub-μs cache hits (≈800 ns, fast path at 0.16 μs, claimed 92× speedup vs KimU) and tens-of-μs misses… | ||
| Xerxes: Extensive Exploration of Scalable Hardware Systems with CXL-Based Simulation... | Yuda An, Shushu Yi, Bo Mao, Qiao Li, Mingzhe Zhang, Diyu Zhou, Ke Zhou, Nong Xiao, Guangyu Sun, Yingwei Luo, Jie Zhang | 11m |
Simulation framework that models CXL 3 port-based routing and device-managed coherence and integrates DRAMSim3 for memory fidelity. Claims calibrated to real… | ||
| Characterizing and Emulating FDP SSDs with WARP | Inho Song, Shoaib Asif Qazi, Javier González, Matias Bjørling, Sam H. Noh, Huaicheng Li | 17m |
Presents WARP, an open FPGA‑based emulator and study that shows FDP SSDs sustain near-1 WAF only when RUH isolation matches object lifetimes… | ||
| ScaleSwap: A Scalable OS Swap System for All-Flash Swap Arrays | Taehwan Ahn, Chanhyeong Yu, Sangjin Lee, Yongseok Son | 14m |
Implements a Linux-kernel swap redesign that assigns per-core swap regions and metadata, adds one-to-one swap model for local allocation, and uses… | ||
| DPAS: A Prompt, Accurate and Safe I/O Completion Method for SSDs | Dongjoo Seo, Jihyeon Jung, Yeohwan Yoon, Ping-Xiang Chen, Yongsoo Joo, Sung-Soo Lim, Nikil Dutt | 14m |
DPAS introduces PAS, which estimates sleep by asking whether the last IO woke early or late using only the two most recent completions, and a per-core state… | ||
| How Soon is Now? Preloading Images for Virtual Disks with ThinkAhead | Xinqi Chen, Yu Zhang, Erci Xu, Changhong Wang, Jifei Yi, Qiuping Wang, Shizhuo Sun, Zhongyu Wang, Haonan Wu, Junping Wu, Hailin Peng, Rong Liu, Yinhu Wang, Jiaji Zhu, Jiesheng Wu, Guangtao Xue, Patrick P. C. Lee | 18m |
Trace analysis of ~160,000 image loads finds lazy loading causes 40% of slow I/Os (>1s), so ThinkAhead preloads blocks using score-based block selection… | ||
| CoFS: A Filesystem for Fast Container Startup | Li Wang, Jinxu Du, Yang Yang, Qingbo Wu, Tao Liu, Haoze Wu | 17m |
CoFS implements an extended FUSE kernel driver plus an image-time metadata layout that stores a minimal perfect hash and a dense metadata array so lookup needs… | ||
| RosenBridge: A Framework for Enabling Express I/O Paths Across the Virtualization... | Shi Qiu, Li Wang, Jianqin Yan, Ruofan Xiong, Leping Yang, Xin Yao, Renhai Chen, Gong Zhang, Dongsheng Li, Jiwu Shu, Yiming Zhang | 13m |
RosenBridge introduces virtio-ndp as a paravirtual device that offloads NDP logic to uBPF in QEMU userspace, hooks into an io_uring backend for async… | ||
| MlsDisk: Trusted Block Storage for TEEs Based on Layered Secure Logging | Erci Xu, Xinyi Yu, Lujia Yin, Xinyuan Luo, Shaowei Song, Qingsong Chen, Shoumeng Yan, Jiwu Shu, Hongliang Tian, Yiming Zhang | 13m |
MlsDisk implements a layered, log-structured virtual block disk that claims CIFC (confidentiality, integrity, freshness, consistency) for TEEs by decomposing… | ||
| SkySync: Accelerating File Synchronization with Collaborative Delta Generation | Zhihao Zhang, Huiba Li, Lu Tang, Guangtao Xue, Jiwu Shu, Yiming Zhang | 14m |
SkySync reuses storage-layer checksums to avoid recomputing chunk digests, implements CRC32 checksum combining to derive misaligned content-defined chunks, and… | ||
| ParaSync: Exploiting Fine-Grained Parallelism for Efficient File Synchronization | Zhihao Zhang, Lu Tang, Huiba Li, Yue Yu, Guangtao Xue, Jiwu Shu, Yiming Zhang | 15m |
ParaSync breaks the CDC sync bottlenecks by (1) multi-threaded chunking that computes CRC32 per sub-chunk and merges checksums via checksum combination instead… | ||
| Discard-Based Garbage Collection for Distributed Log-Structured Storage Systems in... | Runhua Bian, Liqiang Zhang, Jinxin Liu, Jiacheng Zhang, Jianong Zhong, Jiahao Gu, Hao Guo, Zhihong Guo, Yunhao Li, Fenghao Zhang, Jiangkun Zhao, Yangming Chen, Guojun Li, Ruwen Fan, Haijia Shen, Chengyu Dong, Yao Wang, Rui Shi, Jiwu Shu, Youyou Lu | 17m |
ByteStore implements a new Discard-and-Compaction GC called DisCoGC that reclaims large contiguous stale ranges without rewriting valid data, implemented by… | ||
| PolarStore: High-Performance Data Compression for Large-Scale Cloud-Native Databases | Qingda Hu, Xinjun Yang, Feifei Li, Junru Li, Ya Lin, Yuqi Zhou, Yicong Zhu, Junwei Zhang, Rongbiao Xie, Ling Zhou, Bin Wu, Wenchao Zhou | ? |
PolarStore implements dual-layer compression: a software layer compresses 16 KB pages with per-page adaptive selection between LZ4 and ZSTD (default LZ4 at… | ||
| Unleashing Zoned UFS: Cross-Layer Optimizations for Next-Generation Mobile Storage | Jungae Kim, Jaegeuk Kim, Kyu-Jin Cho, Sungjin Park, Jinwoo Kim, Jieun Kim, Iksung Oh, Chul Lee, Bart Van Assche, Daeho Jeong, Konstantin Vyshetsky, Jin-Soo Kim | 14m |
Zone UFS replaces page maps with zone-level mapping (zones are 1,056 MB, 16 KB pages) and implements a device-side GABM buffer manager, a driver/block-layer… | ||
| DOGI: Data Placement with Oracle-Guided Insights for Log-Structured Systems | Jeeyun Kim, Seonggyun Oh, Jungwoo Kim, Jisung Park, Jaeho Kim, Sungjin Lee, Sam H. Noh | 16m |
Presents DOGI, a data-placement system that combines a hybrid heuristic+ML predictor (hot and frozen filters plus an MLP with six features and 10 outputs) with… | ||
| Getting the MOST out of your Storage Hierarchy with Mirror-Optimized Storage Tiering | Kaiwei Tu, Kan Wu, Andrea C. Arpaci-Dusseau, Remzi H. Arpaci-Dusseau | 16m |
Presents Mirror-Optimized Storage Tiering (MOST), which mirrors a small hot set across tiers and routes requests with a dynamic offload ratio optimizer so no… | ||
| Rearchitecting Buffered I/O in the Era of High-Bandwidth SSDs | Yekang Zhan, Tianze Wang, Zheng Peng, Haichuan Hu, Jiahao Wu, Xiangrui Yang, Qiang Cao, Hong Jiang, Jie Yao | 16m |
WSBuffer is a rearchitecture of buffered I/O that separates small unaligned writes from page cache by introducing a scrap buffer, an opportunistic two-stage… | ||
| FailureMiner: A Joint Key Decision Mining Scheme for Practical SSD Failure Prediction... | Shuyang Wang, Yuqi Zhang, Haonan Luo, Kangkang Liu, Gil Kim, JongSung Na, Claude Kim, Geunrok Oh, Kyle Choi, Ni Xue, Xing He | 14m |
Claims an interpretable pipeline that boosts SSD failure precision by ~38.6% and recall by ~80.5% using production telemetry (70M logs, 85 attributes) from… | ||
| DRBoost: Boosting Degraded Read Performance in MSR-Coded Storage Clusters | Xiao Niu, Guangyan Zhang, Zhiyue Li, Sijie Cai | 18m |
DRBoost implements partial-chunk reconstruction, a reconstruction-friendly coding layout, and a fragmentation-free storage layout and integrates into Ceph. A… | ||
| LESS is More for I/O-Efficient Repairs in Erasure-Coded Storage | Keyun Cheng, Guodong Li, Xiaolu Li, Sihuang Hu, Patrick P. C. Lee | 15m |
LESS layers extended sub-stripes built from Reed-Solomon coding to reduce both data read and non‑contiguous seeks during single-block repair, implemented on… | ||
| Advancing Data Integrity in Linux | Anuj Gupta, Christoph Hellwig, Kanchan Joshi, Vikash Kumar, Javier González, Roshan R Nair, Jinyoung Choi | 14m |
Adds a PI offset so the block layer accepts NVMe metadata with PI at the buffer end (merged in Linux 6.9), and exposes a user-space PI API via io_uring ATTR_PI… | ||
| CETOFS: A High-Performance File System with Host-Server Collaboration for Remote Storage | Wenqing Jia, Dejun Jiang, Jin Xiong | 14m |
CETOFS implements a userspace-kernel collaborative architecture for NVMe-over-RDMA remote SSDs, moving the data path into user space and offloads permission… | ||
| Cache-Centric Multi-Resource Allocation for Storage Services | Chenhao Ye, Shawn (Wanxiang) Zhong, Andrea C. Arpaci-Dusseau, Remzi H. Arpaci-Dusseau | 18m |
HARE is a cache-centric allocator that models demand vectors as functions of cache size and jointly optimizes cache plus cache-correlated resources using an… | ||
| Lockify: Understanding Linux Distributed Lock Management Overheads in Shared Storage | Taeyoung Park, Yunjae Jo, Daegyu Han, Beomseok Nam, Jaehyun Hwang | 13m |
Lockify implements a kernel DLM change in Linux 6.6.23 that replaces owner queries during file/directory creation with self-owner notifications and… | ||
| uCache: A Customizable Unikernel-based IO Cache | Ilya Meignan--Masson, Masanori Misono, Viktor Leis, Pramod Bhatotia | 16m |
uCache is a unikernel libOS cache (implemented on OSV) that exposes shared uVFS abstraction and buffer/VMA callbacks so applications can attach replacement… | ||
| UnICom: A Universally High-Performant I/O Completion Mechanism for Modern Computer... | Riwei Pan, Yu Liang, Sam H. Noh, Lei Li, Nan Guan, Tei-Wei Kuo, Chun Jason Xue | 22m |
UnICom implements three kernel mechanisms—TagSched, TagPoll, and SKIP—in Linux to consolidate polling on a dedicated core, retain tagged sleeping tasks on CPU… | ||
| nothing matches | ||