OSDI '25
who it's for
systems researchers, distributed systems engineers, storage engineers, database engineers, machine learning systems engineers, gpu systems engineers, operating systems developers
topics
LLM serving autoscaling; LLM inference at wafer scale; gpu kernel profiling; tensor program superoptimization; RDMA remote allocation; replicated state machine communication; speculative distributed transactions; vector search adaptive indexing; bounded-staleness cache replication; CXL pod distributed database; kernel record replay; encrypted semantic search
| title | author | length |
|---|---|---|
| Accelerating Software Development: The LLM (R)evolution | Emery D. Berger | 44m |
Argues that existing tools get profiler++ behavior with Scalene calling an LLM to suggest optimizations that sometimes yield up to 90x speedups; presents… | ||
| Basilisk: Using Provenance Invariants to Automate Proofs of Undecidable Protocols | Tony Nuda Zhang, Keshav Singh, Tej Chajed, Manos Kapritsos, Bryan Parno | 16m |
Basilisk is a Daphne compiler extension that automatically derives inductive invariants for distributed protocols using Provenance Invariants and an analysis… | ||
| Deriving Semantic Checkers from Tests to Detect Silent Failures in Production... | Chang Lou, Dimas Shidqi Parikesit, Yujin Huang, Zhewen Yang, Senapati Diwangkara, Yuzhuo Jing, Achmad Imam Kistijantoro, Ding Yuan, Suman Nath, Peng Huang | 18m |
Claims an automated pipeline T2C that transforms existing system tests into runtime semantic checkers using static slicing, dynamic tracing, symbolization and… | ||
| Picsou: Enabling Replicated State Machines to Communicate Efficiently | Reginald Frank, Micah Murray, Chawinphat Tankuranand, Junseo Yoo, Ethan Xu, Natacha Crooks, Suyash Gupta, Manos Kapritsos | 16m |
Picsou introduces a formal primitive, Cross-Cluster Consistent Broadcast (C3B), and a practical protocol that uses TCP-like cumulative ACKs plus rotating… | ||
| FineMem: Breaking the Allocation Overhead vs. Memory Waste Dilemma in Fine-Grained... | Xiaoyang Wang, Yongkun Li, Kan Wu, Wenzhe Zhu, Yuqi Li, Yinlong Xu | 19m |
FineMem implements an RDMA-based remote allocator that removes MR registration from the allocation path via per-compute-node pre-registered MRs and isolates… | ||
| To PRI or Not To PRI, That's the question | Yun Wang, Liang Chen, Jie Ji, Xianting Tian, Ben Luo, Zhixiang Wei, Zhibai Huang, Kailiang Xu, Kaihuan Peng, Kaijie Guo, Ning Luo, Guangjian Wang, Shengdong Dai, Yibin Shen, Jiesheng Wu, Zhengwei Qi | 17m |
Software-only dynamic passthrough called VIO prevents device DMA page faults without hardware changes by inserting a shadow available queue and IOPA-snooping… | ||
| Enabling Efficient GPU Communication over Multiple NICs with FuseLink | Zhenghang Ren, Yuxuan Li, Zilong Wang, Xinyang Huang, Wenxue Li, Kaiqiang Xu, Xudong Liao, Yijun Sun, Bowen Liu, Han Tian, Junxue Zhang, Mingfei Wang, Zhizhen Zhong, Guyue Liu, Ying Zhang, Kai Chen | 17m |
FuseLink remaps GPU-side network buffers to relay traffic over high-speed intra-server fabrics (MVLink/MLink) so idle 400G RDMA NICs act as a merged “super… | ||
| Tigon: A Distributed Database for a CXL Pod | Yibo Huang, Haowei Chen, Newton Ni, Yan Sun, Vijay Chidambaram, Dixin Tang, Emmett Witchel | 19m |
Tigon implements a distributed transactional in-memory database for a CXL pod, running on an emulated 8-host setup (40-core machine + CXL 1.1 module). It… | ||
| Mako: Speculative Distributed Transactions with Geo-Replication | Weihai Shen, Yang Cui, Siddhartha Sen, Sebastian Angel, Shuai Mu | 17m |
A transactional key-value store called Mako is built that decouples execution from geo-replication, running speculative 2PC locally and replicating decisions… | ||
| Quake: Adaptive Indexing for Vector Search | Jason Mohoney, Devesh Sarda, Mengze Tang, Shihabur Rahman Chowdhury, Anil Pacaci, Ihab F. Ilyas, Theodoros Rekatsinas, Shivaram Venkataraman | 15m |
Quake implements a multi-level partitioning index with a practical cost model that picks splits/merges by estimated query cost, an online adaptive partition… | ||
| Achieving Low-Latency Graph-Based Vector Search via Aligning Best-First Search... | Hao Guo, Youyou Lu | 12m |
An on-disk system called PipeANN aligns best-first graph search with SSD behavior by issuing speculative, pipelined reads decided from the in-memory candidate… | ||
| Skybridge: Bounded Staleness for Distributed Caches | Robert Lyerly, Scott Pruett, Kevin Doherty, Greg Rogers, Nathan Bronson, John Hugg | 16m |
Skybridge is an out-of-band replication service that indexes write metadata and answers cache staleness queries so caches avoid needless refills. It uses… | ||
| KPerfIR: Towards a Open and Compiler-centric Ecosystem for GPU Kernel Performance... | Yue Guan, Yuanwei Fang, Keren Zhou, Corbin Robeck, Manman Ren, Zhongkai Yu, Yufei Ding, Adnan Aziz | 14m |
KPerfIR is a compiler‑centric infrastructure integrated into Triton that instruments a record op across a multi‑layer IR, adds compiler passes for automatic… | ||
| Mirage: A Multi-Level Superoptimizer for Tensor Programs | Mengdi Wu, Xinhao Cheng, Shengyu Liu, Chunan Shi, Jianan Ji, Man Kit Ao, Praveen Velliengiri, Xupeng Miao, Oded Padon, Zhihao Jia | 16m |
Mirage is a multi-level superoptimizer that generates custom CUDA kernels from tensor programs using a uniform representation across kernel, thread-block, and… | ||
| QiMeng-Xpiler: Transcompiling Tensor Programs for Deep Learning Systems with a... | Shouyang Dong, Yuanbo Wen, Jun Bi, Di Huang, Jiaming Guo, Jianxing Xu, Ruibai Xu, Xinkai Song, Yifan Hao, Ling Li, Xuehai Zhou, Tianshi Chen, Qi Guo, Yunji Chen | 11m |
QiMeng-Xpiler automatically transcompiles tensor programs by combining LLM-assisted transformation passes with SMT-based repair of small faulty snippets, and… | ||
| WaferLLM: Large Language Model Inference at Wafer Scale | Congjie He, Yeqi Huang, Pei Mu, Ziming Miao, Jilong Xue, Lingxiao Ma, Fan Yang, Luo Mai | 17m |
WaferLLM adapts LLM inference to wafer-scale chips using a PLMR model that guides placement and communication, implements MeshGEMM and MeshGEMV operators, and… | ||
| BlitzScale: Fast and Live Large Model Autoscaling with O(1) Host Caching | Dingyan Zhang, Haotian Wang, Yang Liu, Xingda Wei, Yizhou Shan, Rong Chen, Haibo Chen | 14m |
System claims live, fast autoscaling by avoiding full model caching with O(1) host caching and loading parameters over compute fabrics (RDMA) using a… | ||
| Bayesian Code Diffusion for Efficient Automatic Deep Learning Program Optimization | Isu Jeong, Seulki Lee | 17m |
Introduces Bayesian code diffusion that clusters similar subgraphs, picks a prior via cosine similarity of tensor shapes, and propagates loop-splitting… | ||
| Training with Confidence: Catching Silent Errors in Deep Learning Training with... | Yuxuan Jiang, Ziming Zhou, Boyu Xu, Beijie Liu, Runhui Xu, Peng Huang | 17m |
TRAINCHECK automatically infers low‑level training invariants from reference traces and uses selective dynamic instrumentation (monkey‑patching and proxies) to… | ||
| Neutrino: Fine-grained GPU Kernel Profiling via Programmable Probing | Songlin Huang, Chenshu Wu | 15m |
Neutrino inserts instruction-level assembly probing into GPU kernels and uses an eBPF-like programmable probe DSL (Python front end) plus a driver hook that… | ||
| Principles and Methodologies for Serial Performance Optimization | Sujin Park, Mingyu Guan, Xiang Cheng, Taesoo Kim | 15m |
Reviewing 477 OSDI/SOSP papers from the past decade yields a claim that sequential optimizations fall into three principles—remove, replace, reorder—and eight… | ||
| Söze: One Network Telemetry Is All You Need for Per-flow Weighted Bandwidth Allocation... | Weitao Wang, T. S. Eugene Ng | 16m |
Söze uses a single network telemetry signal—the maximum queuing delay from the path (via CIC format)—to drive decentralized per-flow rate updates and realize… | ||
| Decouple and Decompose: Scaling Resource Allocation with DeDe | Zhiying Xu, Minlan Yu, Francis Y. Yan | 16m |
DeDe reformulates large resource-allocation LPs/MIPs by making a matrix copy, applying ADMM to decouple row (resource) and column (demand) constraints, then… | ||
| Quantum Virtual Machines | Runzhou Tao, Hongzheng Zhu, Jason Nieh, Jianan Yao, Ronghui Gu | 19m |
HyperQ builds quantum virtual machines that partition hardware by repeating topology tiles, compiles programs to those smaller virtual back ends, and… | ||
| QOS: Quantum Operating System | Emmanouil Giortamis, Francisco Romão, Nathaniel Tornow, Pramod Bhatotia | 15m |
QOS is a modular quantum OS with an LLVM-like compiler and a runtime (estimator, compatibility-based multi-programming, scheduler) implemented in Python on top… | ||
| Scalio: Scaling up DPU-based JBOF Key-value Store with NVMe-oF Target Offload | Xun Sun, Mingxing Zhang, Yingdi Shan, Kang Chen, Jinlei Jiang, Yongwei Wu, Quan Cheng Laboratory | 15m |
Scalio is a DPU-based JBOF key-value store that bypasses weak ARM cores by using NVMe-oF Target Offload for SSD I/O and one-sided RDMA for cache access, adds… | ||
| Low End-to-End Latency atop a Speculative Shared Log with Fix-Ante Ordering | Shreesha Gopalakrishna Bhat, Tony Hong, Xuhao Luo, Jiyu Hu, Aishwarya Ganesan, Ramnatthan Alagappan | 17m |
Proposes SpecLog, a speculative shared log that predetermines global cuts via fix-ante ordering, implements it as Belfast, and overlaps application compute… | ||
| Understanding Stragglers in Large Model Training Using What-if Analysis | Jinkun Lin, Ziheng Jiang, Zuquan Song, Sida Zhao, Menghan Yu, Zhanghan Wang, Chenyuan Wang, Zuocheng Shi, Xiang Shi, Wei Jia, Zherui Liu, Shuguang Wang, Haibin Lin, Xin Liu, Aurojit Panda, Jinyang Li | 16m |
Five-month ByteDance cluster traces, with jobs up to >5,000 GPUs, drive a counterfactual what-if simulation that replaces computation times with the mean and… | ||
| Fork in the Road: Reflections and Optimizations for Cold Start Latency in Production... | Xiaohu Chai, Tianyu Zhou, Keyang Hu, Jianfeng Tan, Tiwei Bie, Anqi Shen, Dawei Shen, Qi Xing, Shun Song, Tongkai Yang, Le Gao, Feng Yu, Zhengyu He, Dong Du, Yubin Xia, Kang Chen, Yu Chen, Quan Cheng Laboratory | 16m |
Platform traces show 19.5% of requests are cold starts and 18% of invocations finish within 221 ms, exposing three overlooked bottlenecks: control path… | ||
| Kamino: Efficient VM Allocation at Scale with Latency-Driven Cache-Aware Scheduling | David Domingo, Hugo Barbalho, Marco Molinaro, Kuan Liu, Abhisek Pan, David Dion, Thomas Moscibroda, Sudarsun Kannan, Ishai Menache | 17m |
Kamino implements a latency-driven cache-aware scheduler on top of Protein, adding a multiQ design, request classifiers, a latency estimator, and the LAC cache… | ||
| ZEN: Empowering Distributed Training with Sparsity-driven Data Synchronization | Zhuang Wang, Zhaozhuo Xu, Jingyi Xi, Yuke Wang, Anshumali Shrivastava, T. S. Eugene Ng | 17m |
Claims a near‑optimal sparse-gradient synchronization that cuts communication up to 5.09x and boosts training throughput up to 2.48x vs prior work, with… | ||
| Extending Applications Safely and Efficiently | Yusheng Zheng, Tong Yu, Yiwei Yang, Yanpeng Hu, Xiaozheng Lai, Dan Williams, Andi Quinn | 15m |
EIM models each extension feature as a named resource and splits work into a developer-made capability map and a deployment-time YAML that grants a minimal… | ||
| Tintin: A Unified Hardware Performance Profiling Infrastructure to Uncover and Manage... | Ao Li, Marion Sudvarg, Zihan Li, Sanjoy Baruah, Chris Gill, Ning Zhang | 20m |
Tintin is a Linux-kernel PMU subsystem with three modules—monitor, scheduler, manager—that quantifies multiplexing error at runtime, schedules events to… | ||
| Building Bridges: Safe Interactions with Foreign Languages through Omniglot | Leon Schuermann, Jack Toubes, Tyler Potyondy, Pat Pannuto, Mae Milano, Amit Levy | 17m |
Omniglot enforces Rust soundness when calling unmodified C libraries by interposing on the FFI, validating values at call boundaries and tracking foreign… | ||
| KRR: Efficient and Scalable Kernel Record Replay | Tianren Zhang, Sishuai Gong, Pedro Fonseca | 16m |
KRR implements a split-recorder with one recorder in the hypervisor and one inside the guest kernel, recording only inputs that enter the kernel and writing a… | ||
| Disentangling the Dual Role of NIC Receive Rings | Boris Pismenny, Adam Morrison, Dan Tsafrir | 18m |
Claim: rxBisect splits each Rx ring into two independent rings so allocation and packet delivery decouple, implemented with software NIC emulation and… | ||
| Deterministic Client: Enforcing Determinism on Untrusted Machine Code | Zachary Yedidia, Geoffrey Ramseyer, David Mazières | 17m |
Deterministic Client (DeCl) enforces determinism for x86-64 and Arm64 machine code by rewriting LLVM output and verifying machine-code verifier constraints… | ||
| XSched: Preemptive Scheduling for Diverse XPUs | Weihang Shen, Mingcong Han, Jialong Liu, Rong Chen, Haibo Chen | 15m |
XSched implements a preemptible command-queue called XQ and a multi-level hardware model (three levels: pending, inflight, running) to bring preemption to… | ||
| OS Rendering Service Made Parallel with Out-of-Order Execution and In-Order Commit | Yuanpei Wu, Dong Du, Chao Xu, Yubin Xia, Yang Yu, Ming Fu, Binyu Zang, Haibo Chen | 16m |
Spars generates self-contained rendering tasks via in-order preparation, executes them across worker threads with out-of-order execution, and enforces drawing… | ||
| EMT: An OS Framework for New Memory Translation Architectures | Siyuan Chai, Jiyuan Zhang, Jongyul Kim, Alan Wang, Fan Chung, Jovan Stojkovic, Weiwei Jia, Dimitrios Skarlatos, Josep Torrellas, Tianyin Xu | 17m |
EMT is an OS framework that ports Linux memory management onto a hardware‑neutral MMU interface, enabling radix, hash, and direct‑map translation schemes with… | ||
| Tiered Memory Management Beyond Hotness | Jinshu Liu, Hamid Hadian, Hanchen Xu, Huaicheng Li | 16m |
Introduces amortized offcore latency (AOL), a metric that divides off-chip latency by observed MLP to predict performance impact; builds an offline… | ||
| NanoFlow: Towards Optimal Large Language Model Serving Throughput | Kan Zhu, Yufei Gao, Yilong Zhao, Liangyu Zhao, Gefei Zuo, Yile Gu, Dedong Xie, Tian Tang, Qinyu Xu, Zihao Ye, Keisuke Kamahori, Chien-Yu Lin, Ziren Wang, Stephanie Wang, Arvind Krishnamurthy, Baris Kasikci | 16m |
NanoFlow splits request batches into nano-batches and duplicates per-batch operations to overlap compute, memory (KV cache), and network work within a single… | ||
| PipeThreader: Software-Defined Pipelining for Efficient DNN Execution | Yu Cheng, Lei Wang, Yining Shi, Yuqing Xia, Lingxiao Ma, Jilong Xue, Yang Wang, Zhiwen Mo, Feiyang Chen, Fan Yang, Mao Yang, Zhi Yang | 15m |
A compiler shifts pipeline scheduling from hardware to software, exposing tile-level tasks and heterogeneous units to generate efficient DNN pipelines. It… | ||
| WLB-LLM: Workload-Balanced 4D Parallelism for Large Language Model Training | Zheng Wang, Anna Cai, Xinfeng Xie, Zaifeng Pan, Yue Guan, Weiwei Chu, Jie Wang, Shikai Li, Jianyu Huang, Chris Cai, Yuchen Hao, Yufei Ding | 16m |
Analysis finds up to 1.4× GPU compute skew in Llama 3 400B traces and pins sources to pipeline-level packing and context-level sharding. The system implements… | ||
| DecDEC: A Systems Approach to Advancing Low-Bit LLM Quantization | Yeonhong Park, Jake Hyun, Hojoon Kim, Jae W. Lee | 16m |
DecDEC stores the weight residuals on CPU and dynamically fetches only salient channels (top-K by activation outliers) via zero-copy to correct low-bit errors… | ||
| Stripeless Data Placement for Erasure-Coded In-Memory Storage | Jian Gao, Jiwu Shu, Bin Yan, Yuhao Zhang, Keji Huang | 11m |
Stripeless scheme replaces stripe coordination by having each primary replicate to K backups and each backup XOR-encodes the K replicas into parity, using a… | ||
| Fast and Synchronous Crash Consistency with Metadata Write-Once File System | Yanqi Pan, Wen Xia, Yifeng Zhang, Xiangyu Zou, Hao Huang, Zhenhua Li, Chentao Wu | 15m |
Build: a Linux kernel prototype called WOLVES implementing a metadata write-once file system that turns each file operation into a single checksum-protected… | ||
| Decentralized, Epoch-based F2FS Journaling with Fine-grained Crash Recovery | Yaotian Cui, Zhiqi Wang, Renhai Chen, Zili Shao | 16m |
Claims a working prototype that journals only metadata, embeds logs into inodes, and uses an epoch-based control/data decoupling so new transactions do not… | ||
| PoWER Never Corrupts: Tool-Agnostic Verification of Crash Consistency and Corruption... | Hayley LeBlanc, Jacob R. Lorch, Chris Hawblitzel, Cheng Huang, Yiheng Tao, Nickolai Zeldovich, Vijay Chidambaram | 18m |
Claims tool-agnostic verification of crash consistency via PoWER (Preconditions on Writes Enforcing Recoverability), implements an explicit model of media… | ||
| Okapi: Decoupling Data Striping and Redundancy Grouping in Cluster File Systems | Sanjith Athlur, Timothy Kim, Saurabh Kadekodi, Francisco Maturana, Xavier Ramos, Arif Merchant, K.V. Rashmi, Gregory R. Ganger | 17m |
Okapi decouples stripes and redundancy groups, letting stripe width be tuned for IO patterns while group width is tuned for reliability and cost; prototype… | ||
| Weave: Efficient and Expressive Oblivious Analytics at Scale | Mahdi Soleimani, Grace Jia, Anurag Khandelwal | 15m |
Weave implements an oblivious MapReduce pipeline that adds 1% sampling to build a histogram, inserts intermediate “vavers” to perform a two‑stage random then… | ||
| Compass: Encrypted Semantic Search with High Accuracy | Jinhao Zhu, Liana Patel, Matei Zaharia, Raluca Ada Popa | 16m |
Compass implements encrypted semantic search by running HNSW-style graph traversal on encrypted embeddings with a client-side retriever and ORAM-backed… | ||
| Paralegal: Practical Static Analysis for Privacy Bugs | Justus Adam, Carolyn Zech, Livia Zhu, Sreshtaa Rajesh, Nathan Harbison, Mithi Jethwa, Will Crichton, Shriram Krishnamurthi, Malte Schwarzkopf | 20m |
Paralegal is a static analysis tool for Rust that checks formal privacy policies over a marker-annotated Program Dependence Graph, using Rust’s ownership types… | ||
| MettEagle: Costs and Benefits of Implementing Containers on Microkernels | Till Miemietz, Viktor Reusch, Matthias Hille, Lars Wrenger, Jana Eisoldt, Jan Klötzke, Max Kurze, Adam Lackorzynski, Michael Roitzsch, Hermann Härtig | 16m |
Prototype MettEagle runs on L4Re with only small, three‑digit LOC changes to the base system, adds an in‑memory filesystem and UDP/IP stack, and ports Python… | ||
| nothing matches | ||