20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26)
2026-07-13..2026-07-15
Seattle, United States
who it's for
systems researchers, cloud infrastructure engineers, ml systems engineers, operating system developers, storage and file system engineers, database engineers, security engineers, browser engine developers
topics
CXL memory disaggregation; host-backed GPU KV caching; zero-copy GPU CPU KV access; elastic channel NUMA memory; memory telemetry and observability; scalable LLM inference serving; batch inference architectures; disaggregated ML training; kernel performance tunability; multi-component file system coordination; userspace network runtime
| title | author | length |
|---|---|---|
| ECHO: Efficient KV Cache Offloading with Lossless Prefetching for Serving Native... | Guangda Liu, Wenhao Chen, Chengwei Li, Zhenyu Ning, Jing Lin, Yiwu Yao, Quan Chen, Shixuan Sun, Jieru Zhao, Minyi Guo | 15m |
ECHO offloads KV caches to host memory, expanding the effective GPU K pool from ~330K to 1.8M tokens and delivering up to 2.1× higher generation throughput on… | ||
| No Buffer, No Bottleneck: Efficient Zero-Copy KV Cache Offloading for Long-Context LLMs | Shutian Luo, Haiying Shen | 16m |
DirectKV implements a true zero-copy KV cache by letting CUDA kernels directly access CPU-resident KV blocks over NVLink-C2C on GH200/GB200, eliminating the… | ||
| Break On Through to the Other Side: Pooling Memory Elastically with RamRyder | Yanbo Zhou, Erci Xu, Dongjoo Seo, Adam Manzanares, Steven Swanson | 17m |
RamRyder is a software-defined elastic memory system that maps guest physical pages to specific DRAM channels (obtained via BIOS/UEFI partitioning and reverse… | ||
| MAC: Metadata Acceleration for Sustainable Performance in Big-Data Systems with CXL DRAM | Dusol Lee, Yan Sun, Houxiang Ji, Vinit Gupta, Austin Antony Cruz, Inhyuk Choi, Nam Sung Kim, Jihong Kim | 15m |
MAC targets Linux metadata spillover into CXL DRAM and places metadata in CXL DRAM while offloads page reclamation to NMP accelerators. With a 1:4 DDR:CXL… | ||
| Finding NEMO: Nimble and Expressive Memory Observability | Shihang Li, Matthew Giordano, Tushar Garg, Rohan Kadekodi, Daniel S. Berger, Baris Kasikci, Thomas E. Anderson, Simon Peter | 15m |
NEMO implements a programmable three‑stage telemetry pipeline in the memory controller and prototypes it on an FPGA-based CXL memory controller (Altera Agel 7… | ||
| USEC: A User-Requirement-Driven Mandatory Access Control Framework for Operating Systems | Yu Jiang, Wenhuan Liu, Fuchen Ma, Yuheng Shen, Yuanliang Chen, Lei Zhang, He Li, Quan Zhang, Chijin Zhou | 19m |
Implements a Linux security extension that anchors policies at protected files, sockets and device nodes instead of process domains, compiles those templates… | ||
| Mohabi: Disaggregating and Sandboxing the Firefox JavaScript Engine | Abhishek Sharma, Anand Balaji, Zachary Yedidia, Anthony Du, Taehyun Noh, Iain Ireland, Jan de Mooij, Matthew Gaudet, Tal Garfinkel, Deian Stefan, Hovav Shacham, Shravan Ravi Narayan | 18m |
Mohabi disaggregates SpiderMonkey from Firefox, adds wrapper types (e.g., split rooted wrapper, sandbox heap/stack/callbacks), and applies an external SFI… | ||
| Ichnaea: A Framework for Precise Tracking of Memory Objects | Samad Haque, Sibin Mohan, Aaron Paulos, Partha Pal | 14m |
A framework that uses memory protection keys and a user-space page-fault handler to log per-access context (accessor, value, timestamp, call stack) and… | ||
| Extracting Database Access-Control Policies from Web Applications | Wen Zhang, Dev Bali, Jamison Kerney, Aurojit Panda, Scott Shenker | 18m |
Ote concretely extracts implicit DB access-control policies from Ruby on Rails apps by running concolic execution to produce traces of SQL and the conditions… | ||
| Tessera: A Holistic Pipeline Parallelism Framework for Trillion-Parameter Heterogeneous | Weifang Hu, Langshi Chen, Man Yuan, Youyang Yao, Xiulong Yuan, Li Tian, Yong Li, Wei Lin, Xuanhua Shi, Zhengping Qian, Jingren Zhou | 17m |
Tessera is a pipeline-parallel framework that profiles fine-grained, per-pair overlap schedules and uses those measured post-overlap costs to cut stages, then… | ||
| Hetu v2: A General and Scalable Deep Learning System with Hierarchical and Heterogeneous | Haoyang Li, Fangcheng Fu, Hao Ge, Sheng Lin, Xuanyu Wang, Jiawen Niu, Yuming Zhou, Xupeng Miao, Bin Cui | 15m |
Hetu v2 implements HSPMD, an extension of SPMD that adds a top-tier subgroup sharding and asymmetric communication primitives so tensors can be unevenly… | ||
| Cocoon: A System Architecture for Differentially Private Training with Correlated Noises | Donghwan Kim, Xin Gu, Jinho Baek, Timothy Lo, Younghoon Min, Kwangsik Shin, Jongryool Kim, Jongse Park, Kiwan Maeng | 15m |
Cocoon presents a system that stores and computes correlated noises across GPU, CPU, and a CXL near‑memory processor to avoid the 14× slowdowns observed when a… | ||
| ValScope: Value-Semantics-Aware Metamorphic Testing for Detecting Logical Bugs in DBMSs | Li Lin, Liehang Chen, Rongxin Wu | 18m |
ValScope implements a unified SQL query approximation model and an approximation propagation analysis to reason how local mutations (aggregation and expression… | ||
| When Sampling Lies: Trustworthy Performance Profiling for Flat Workloads with Blink | Rishikesh Devsot, ChenXing Yang, Yi Fan Yu, Prabhdeep Singh Soni, Afshin Arefi, Bryan Chan, Reza Azimi, Ding Yuan | 14m |
Blink instruments every function at entry and exit, reads PMU counters at those points to compute precise deltas, and uses self-patching to replace tracing… | ||
| Breaking the Reward Barrier: Accelerating Tree-of-Thought Reasoning via Speculative... | Shuzhang Zhong, Haochen Huang, Shengxuan Qiu, Pengfei Zuo, Runsheng Wang, Meng Li | 14m |
Proposes SPEX, a speculative-exploration system implemented on SGLang that predicts and expands likely Tree-of-Thought branches before reward feedback, using… | ||
| Controlling Opaque-Component Effects with Semisolates and Try | Evangelos Lamprou, Di Jin, Grigoris Ntousakis, Georgios Liargkovas, Calvin Eng, Konstantinos Kallas, Michael Greenberg, Nikos Vasilakis | 12m |
Introduces semisolates and the unprivileged command try that capture, inspect, stack, and selectively apply filesystem effects from opaque components while… | ||
| SBB: Eliminating Centralized Bottlenecks in Userspace Network Runtime | Kang Hu, Shuqi Dong, Chuandong Li, Ran Yi, Zonghao Zhang, Yiming Yao, Bo An, Jie Zhang, Xiaolin Wang, Yingwei Luo, Zhenlin Wang, Diyu Zhou | 17m |
SBB is a decentralized userspace network runtime that removes timer, monitor, and dispatcher bottlenecks by delivering user-level timer interrupts (via LAPIC)… | ||
| Rakaia: Scalable In-Kernel Scheduling for TCP-Based RPCs | Rui Yang, Konstantinos Prasopoulos, Edouard Bugnion | 13m |
Builds a message-oriented API inside the Linux kernel by parsing TCP receive streams early, implemented as a 3000-line kernel module plus a 60-line… | ||
| Xkernel: Principled Performance Tunability of Operating System Kernels | Zhongjie Chen, Wentao Zhang, Yulong Tang, Ran Shu, Fengyuan Ren, Tianyin Xu, Jing Liu | 16m |
Xkernel converts frozen Linux perf-consts into safe, in-situ knobs by synthesizing runtime handlers via Scoped Indirect Execution; it finds constants with… | ||
| Murakkab: Resource-Efficient Agentic Workflow Orchestration in Cloud Platforms | Gohar Irfan Chaudhry, Esha Choukse, Haoran Qiu, Íñigo Goiri, Rodrigo Fonseca, Adam Belay, Ricardo Bianchini | 16m |
Murakkab exposes agentic workflows as a declarative workflow spec and uses an LLM to pick agents then a profile-guided optimizer plus adaptive runtime to pick… | ||
| StriaTrace: Efficient Tracing and Diagnosis for Online LLM Inference | Haonan Wu, Yanqing Chen, Kun Qian, Xue Li, Jingbo Xu, Erci Xu, Ennan Zhai, Wenyuan Yu, Guangtao Xue, Jingren Zhou | 12m |
StriaTrace implements host and device collectors that record only 13 synchronization spans and kernel boundaries plus memory transfers, with a pass-by sampler… | ||
| Diagnosing Performance Issues in Application-Defined Resources | Yigong Hu, You-Liang Huang, Haodong Zheng, Yicheng Liu, Dedong Xie, Baris Kasikci | 18m |
gigiprofiler is a C/C++ tool built on LLVM that combines LLM-based semantic inference (GPT‑4) with static validation and runtime instrumentation to expose… | ||
| hS: Speculative Script Reordering at Subprocess Granularity | Georgios Liargkovas, Di Jin, Tianyu (Ezri) Zhu, Dan Liu, A. Bolun Thompson, Anirudh Narsipur, Seong Heon Jung, Siddhartha Prasad, Diomidis Spinellis, Michael Greenberg, Konstantinos Kallas, Nikos Vasilakis | 16m |
hS is a system that speculatively executes shell commands out of order by running command instances in sandboxed layers and tracing system calls to detect real… | ||
| Incr: Faster Re-Execution via Bolt-On Incrementalization | Yizheng Xie, Evangelos Lamprou, Jerry Xia, Nikos Vasilakis | 14m |
Incr is a bolt-on system that incrementalizes unmodified shell programs by tracing command effects, running each command in a semi-isolated overlay, and… | ||
| Aletheia: Automated Detection of Data Integrity Violations in Microservices | Mafalda Sofia Ferreira, João Ferreira Loff, Joao Garcia, Rodrigo Rodrigues | 16m |
Formalizes integrity violations with the Entity-Relationship model and relational algebra and implements a static analyzer, Aletheia, that uses static single… | ||
| Arctic: A Practical Lock-Free Adaptive Radix Tree | Newton Ni, Nicolas Garza, Jenny Stinehour, Michael Goppert, Michal Friedman, Emmett Witchel | 17m |
Arctic is a lock‑free adaptive radix tree that uses a freezing‑based coordination protocol and edge‑in‑parent layout to perform compress/replace operations… | ||
| RLinf: Flexible and Efficient Large-Scale Reinforcement Learning... | Chao Yu, Yuanqing Wang, Zhen Guo, Hao Lin, Si Xu, Hongzhi Zang, Quanlu Zhang, Yongji Wu, Chunyang Zhu, Junhao Hu, Zixiao Huang, Mingjie Wei, Yuqing Xie, Ke Yang, Bo Dai, Zhexuan Xu, Jiakun Du, Xiangyuan Wang, Xu Fu, Letong Shi, Zhihao Liu, Kang Chen | 17m |
RLinf is a distributed RL training system that implements macro-to-micro flow transformation to break high-level workflows into optimized execution flows; it… | ||
| DynaRL: Flexible and Dynamic Scheduling of Large-Scale Reinforcement Learning Training | Yuanqing Wang, Hao Lin, Junhao Hu, Chunyang Zhu, Quanlu Zhang, Zhen Guo, Yuchen Zhang, Xu Fu, Si Xu, Bo Dai, Zixiao Huang, Chao Yu, Boxun Li, Guohao Dai, Zhi Yang, Yu Wang | 19m |
DynaRL claims a system that reclaims idle GPUs and reallocates them online using a dynamic hypergraph + a unified interrupt and migrate APIs, with three… | ||
| RollArt: Disaggregated Multi-Task Agentic RL Training at Scale | Wei Gao, Yuheng Zhao, Tianyuan Wu, Shaopan Xiong, Weixun Wang, Dakai An, Lunxi Cao, Dilxat Muhtar, Zichen Liu, Haizhou Zhao, Ju Huang, Siran Yang, Yongbin Li, Wenbo Su, Jiamang Wang, Lin Qu, Bo Zheng, Wei Wang | 15m |
RollArt disaggregates multi-task agentic RL by mapping prefill tasks to compute-optimized GPUs, decode tasks to bandwidth-optimized GPUs, and environments to… | ||
| Seer: Online Context Learning for Fast Synchronous LLM Reinforcement Learning | Ruoyu Qin, Weiran He, Weixiao Huang, Yangkun Zhang, Yikai Zhao, Bo Pang, Xinran Xu, Yingdi Shan, Yongwei Wu, Mingxing Zhang | 19m |
Seer claims synchronous on‑policy RL rollouts speed up by up to 2.04× and cuts long‑tail latency 72–94% by exploiting intra‑group similarities in length and… | ||
| Harvesting Sub-Microsecond CXL Memory Stalls with LiteSwitch | Nanqinqin Li, Yuhong Zhong, Asaf Cidon, Michael J. Freedman | 18m |
LiteSwitch is a hardware–software co‑design that harvests sub‑microsecond CXL memory stalls by detecting CXL‑bound loads on the memory path and redirecting… | ||
| Duhu: Shared Disaggregated Memory for Distributed Data Processing Frameworks | Qiutong Men, Tao Wang, Jongryool Kim, Hane (Stella) Yie, Emmanuel Amaro, Marcos K. Aguilera, Aurojit Panda | 17m |
A system that replaces per-node object copying with a segmented, owner-assigned SDM layout and an in‑SDM ring‑buffer RPC channel, implemented on a prototype… | ||
| Espresso: Constructing Cost-Efficient CXL JBOF via Inter-SSD Computing Resource Sharing | Shushu Yi, Yuda An, Li Peng, Xiurui Pan, Qiao Li, Jieming Yin, Guangyan Zhang, Wenfei Wu, Chenxi Wang, Diyu Zhou, Zhenlin Wang, Xiaolin Wang, Yingwei Luo, Ke Zhou, Jie Zhang | 16m |
Espresso builds a JBOF that disaggregates SSDs into disaggregated compute/data ends and uses the cache-coherent CXL fabric to harvest idle onboard ARM cores… | ||
| Accelerating Confidential Databases with Crypto-Free Mappings | Wenxuan Huang, Zhanbo Wang, Mingyu Li | 17m |
ZENO replaces per-field encryption with crypto-free mappings that store compact integer FIDs and a trusted mapping store inside the TEE, performing map lookups… | ||
| Osprey: Transparent and Efficient Virtual Memory for Secure Computation | Yicheng Liu, Alice Yeh, Harry Xu, Raluca Ada Popa, Sam Kumar | 15m |
Osprey provides a runtime virtual memory layer for secure computation that pages encrypted data to SSDs by forking a speculative process and a programmed… | ||
| μUSB: Practical and Safe USB Driver Reuse for Arm TrustZone | Xuankai Zhang, Sijin Li, Pei Meng, Meng Wang, Yongzhao Zhang, Ting Chen, Xiaosong Zhang, Liwei Guo | 18m |
μUSB generates TrustZone USB drivers by record, lift, replay of real Linux driver interactions, using a lightweight mutational recorder plus symbolic… | ||
| BatchGen: An Architecture for Scalable and Efficient Batch Inference | Tairan Xu, Leyang Xue, Zhan Lu, Jinfu Deng, Hongyang Xiao, Yinsicheng Jiang, Congjie He, Matej Sandor, Le Xu, Luo Mai | 20m |
An execution architecture represents every input sequence as a fine-grained event-driven coroutine and exposes sequence coroutine compute model primitives to… | ||
| Kareus: Joint Reduction of Dynamic and Static Energy in Large Model Training | Ruofan Wu, Jae-Won Chung, Mosharaf Chowdhury | 15m |
Kareus jointly optimizes GPU frequency, SM allocation and launch timing, and kernel scheduling using multi-objective Bayesian optimization per partition to… | ||
| SPADE: Signal-Aware DAG Scheduling and Dynamic Provisioning for Data Processing Clusters | Adam Lechowicz, Rohan Shenoy, Noman Bashir, Mohammad H. Hajiesmaili, Adam Wierman, Christina Delimitrou | 15m |
SPADE jointly schedules DAG tasks and provisions resources using a relative importance metric derived from Decima/Graphene scores and an exponential threshold… | ||
| Bodega: Localized Linearizable Reads at Anywhere Anytime via Roster Leases | Guanzhou Hu, Andrea C. Arpaci-Dusseau, Remzi H. Arpaci-Dusseau | 23m |
Bodega introduces roster leases: an all-to-all leasing layer that decouples lease participants from the protected metadata (the roster) so any responder… | ||
| Equal Opportunity: A Correctness Condition for Ordered Consensus | Yunhao Zhang, Haobin Ni, Soumya Basu, Shir Cohen, Maofan Yin, Lorenzo Alvisi, Robbert van Renesse, Qi Chen, Lidong Zhou | 18m |
Defines a formal fairness condition called ordering equality (and a separation parameter delta) and shows randomness is necessary and sufficient to bound bias… | ||
| Jetpack: Consensus Made Generally Fast | Ze Tang, Zihao Zhang, Weihai Shen, Jicheng Shi, Shuai Mu | 19m |
Jetpack is a plugin framework that adds a 1-RTT fast path to existing leader-based consensus protocols with minimal modifications, running fast and original… | ||
| Ambulance: Saving BFT through Racing | Neil Giridharan, Shubham Mishra, Lorenzo Alvisi, Natacha Crooks, Benjamin Marsh, Hein Meling, Kartik Nayak, Grzegorz Prusak | 16m |
Ambulance replaces timeouts with protocol-rigged races where the leader runs a two-step path and others a three-step path so the leader has a one-step head… | ||
| OpGuard: Bitwise Alignment for Precise and General Debugging of Production LLM Training | Ziming Zhou, Yinjie Zhao, Hang Zhu, Wenxiao Wang, Zhihao Bai, Yun Zhang, Shuguang Wang, Haibin Lin, Peng Huang | 17m |
Claims a practical system, OpGuard, that uses bitwise alignment—lightweight per-op fingerprints plus a schedule-tolerant mapper that finds the longest prefix… | ||
| Oxbow: A Coordinated Architecture for Multi-Component File Systems | Jongyul Kim, Jaehwan Lee, Inhoe Koo, Peizhe Liu, Jiyuan Zhang, Junho Ahn, Tianyin Xu, Youngjin Kwon | 18m |
Oxbow builds a three-part file system (user daemon, kernel ufs, device server) that routes reads through the kernel and sends writes directly to a… | ||
| RoCE BALBOA: Service-Enhanced RDMA Offload Engine for Data Center SmartNICs | Maximilian Jakob Heer, Benjamin Ramhorst, Yu Zhu, Luhao Liu, Zhiyi Hu, Jonas Dann, Gustavo Alonso | 18m |
An open-source RDMA offload engine for FPGAs that implements RoCEv2 and sustains 100 Gbps line rate on an Alveo U55C, matching commercial RNICs while keeping… | ||
| DPA-Store: An Ordered Network Data Path Key-Value Store | Frederic Schimmelpfennig, Jan Sass, Reza Salkhordeh, Martin Kröning, Stefan Lankes, André Brinkmann | 16m |
DPA-Store runs a lock-free learned index inside the BlueField-3 DPA memory, fetches values via host DMA to a host-side tree replica, and uses batched writes… | ||
| FARLock: Asymmetric RDMA Locking Made Fair | Yuehao Hu, Jiatang Zhou, Tianzheng Wang, Keval Vora | 18m |
FARLock implements ticket and MCS-style handover plus separate local and remote queues to avoid RDMA loopback while enforcing strict arrival-order grants; it… | ||
| When DDIO Meets Page Coloring: Revisiting DDIO Performance with Sepia | Changwoo Song, Sanghyun Kim, Jinhyeok Oh, Qizhe Cai, Joonsung Kim, Jaehyun Hwang | 16m |
Revisits DDIO’s poor packet-path throughput and shows conflict misses, not just reserved-way capacity, shrink usable LLC. Implements Sepia, a color-aware page… | ||
| Efficient GPU-Centric Evolving Graph Processing at Scale | Yunmo Zhang, Jiacheng Huang, Xizhe Yin, Junqiao Qiu, Hong Xu, Chun Jason Xue | 15m |
POEGA is a GPU-centric framework that uses proxy graphs to keep a compact in-memory approximation, runs a fused kernel to process many snapshots concurrently… | ||
| The Clustering Strikes Back: Building Cost-Effective and High-Performance ANNS at Scale | Yuchen Huang, Baiteng Ma, Yiping Sun, Yang Shi, Xiao Chen, Xiaocheng Zhong, Zhiyong Wang, Yao Hu, Erci Xu, Chuliang Weng | 16m |
Helmsman replaces memory-heavy HNSW with a clustering-based ANNS that uses SPDK userspace storage, a leveling-learned pruning model to pick how many clusters… | ||
| Timelock Drive: Isolated Time-Based Defense for Storage Systems | Jonah Rosenblum, Juechu Dong, Peter Chen, Satish Narayanasamy | 15m |
Timelock Drive adds a time lock command to a storage device that makes physical disk blocks immutable for a set duration, forcing attackers to wait. Metadata… | ||
| Inside Out: A Paradigm Shift in VM Introspection | Dufy Teguia, Louis Duval, Teo Pisenti, Kahina Lazri, Daniel Hagimont, Thomas Pasquier, Renaud Lachaize, Alain Tchana | 14m |
GOODKIT runs observation agents as side VMs colocated with the VMM (built on Firecracker KVM), giving native-speed access to target memory while preserving… | ||
| vBOIDs: Taming Chaos via Coarse-Grained Scheduling Abstraction for Containers | Kaesi Manakkal, Nathan Daughety, Yu Sun, Marcus Pendleton, Hui Lu | 15m |
Introduces a kernel patch (≈2,000 LOC on Linux 6.15) that groups container threads into lightweight placement units called BOID abstraction and coordinates a… | ||
| Efficient LLM Serving on Commodity GPU Clusters with Data-Reduced Cross-Instance... | Jiangsu Du, Hongbin Zhang, Taosheng Wei, Zhenyi Zheng, Jiazhi Jiang, Kaiyi Wu, Zhiguang Chen, Yutong Lu | 19m |
EcoServe is a serving system that implements partially disaggregated (PaDG) scheduling to avoid KV-cache transfers while reducing prefill–decode interference… | ||
| Revisiting Pipeline Parallelism for LLM Serving | Soonjae Hwang, Jeongseob Ahn | 16m |
Revisits pipeline parallelism with an implementation on SGLang that runs on four NVIDIA A100 40GB GPUs and evaluates Qwen2.5 32B and 14B. Introduces greedy and… | ||
| OpenTela: Unifying Decentralized Computing Resources for Heterogeneous LLM Serving (Oper… | Xiaozhe Yao, Youhe Jiang, Ilia Badanin, Qinghao Hu, Robert Matthew Smith, Binhang Yuan, Imanol Schlag, Eiko Yoneki, Ana Klimovic | 19m |
A user-space overlay implements CRDT-based gossip and NAT-punching/relay transport to federate Slurm, Kubernetes and bare-metal serving engines without root or… | ||
| Kairox: Adaptive GPU-CPU Hybrid LLM Inference via Online Neuron Balancing | Yapeng Jiang, Minghao Gan, Zicong Hong, Wuhui Chen, Junyuan Liang, Yue Yu, Meng Guo, Zibin Zheng | 14m |
Kairox is a GPU–CPU hybrid inference system that dynamically moves FFN neuron groups at runtime to trade CPU compute for PCIe transfers, implemented in ~6,000… | ||
| ADAngel: Accelerating Arbitrary-Precision Quantized LLMs with Adaptive Computing Mapping | Yao Liu, Wenjie Wang, Yifei Feng, Bo Peng, Jianguo Yao, Haibing Guan | 17m |
ADAngel delivers a workload-adaptive mpGEMM engine: it implements a unified DPR model to generate padding, bitwise and a new split strategy (which cuts… | ||
| TileLoom: Automatic Dataflow Planning for Tile-Based Languages on Spatial Dataflow Accel… | Wei Li, Zhenyu Bai, Heru Wang, Pranav Dangi, Zhiqiang Zhang, Cheng Tan, Huiying Lan, Weng-Fai Wong, Tulika Mitra | 22m |
TileLoom is an MLIR-based end-to-end compiler that lowers tile-based DSLs (Halide and Triton) to spatial dataflow accelerators by mapping tile grids across a… | ||
| GraCE: Unlocking CUDA Graphs with Compiler Support for ML Workloads | Abhishek Ghosh, Ajay Nayak, Ashish Panwar, Arkaprava Basu | 15m |
GraCE is a compiler framework built atop PyTorch2 that unlocks CUDA Graphs for 25 ML workloads (TorchBench, HuggingFace, TIMM) and requires no code changes. It… | ||
| Continuation-Centric Computing with Arca | Akshay Srivatsan, Yuhan Deng, Katherine Mohr, Emma Sudo, Sebastian Ingino, Francis Chua, Keith Winstein | 21m |
Arca is a continuation-centric OS that provides kernel-level continuation capture via a new syscall pair get_current_continuation and… | ||
| Rethinking Process Snapshots for Near-Warm Serverless Cold Starts | Ben Holmes, Baltasar Dinis, Lana Honcharuk, Adam Belay, Joshua Fried | 14m |
Implements a kernel-side restore path called Spice that adds a new mapping type and file format: SHELF format plus a spliceVMA kernel primitive and a re-exec… | ||
| TrainMover: An Interruption-Resilient Runtime for ML Training | ChonLam Lao, Jiaqi Gao, Jiamin Cao, Zhipeng Zhang, Pengcheng Zhang, Jiangfei Duan, Zhilong Zheng, Yu Guan, Yichi Xu, Yong Li, Zhengping Qian, Aditya Akella, Minlan Yu, Ennan Zhai, Dennis Cai, Jingren Zhou | 17m |
TrainMover is an ML training runtime that prepares standby GPUs ahead of interruptions by pre-recording output tensors and running communication-free sandboxed… | ||
| Nixie: Efficient, Transparent Temporal Multiplexing for Consumer GPUs | Yechen Xu, Yifei Wang, Nathanael Ren, Yiran Chen, Danyang Zhuo | 15m |
Nixie is a system service that temporally multiplexes consumer GPUs by inserting a transparent CUDA shim per application and a centralized daemon that… | ||
| μShell: A Microkernel-based FPGA Shell Architecture | Jiyang Chen, Anubhav Panda, Harshavardhan Unnibhavi, Atsushi Koshiba, Pramod Bhatotia | 15m |
MicroShell implements a microkernel-style FPGA shell with a hardware capability enforcement unit, a dynamic data interconnect, and a software capability… | ||
| PIMS: Fleet-Wide Datacenter Maintenance with Minimal Capacity Buffer and Predictable Lat… | Benjamin Leonhardi, Evangelia Kalyvianaki, Yang Wang, Abdelrahman Adam, Agshin Nabiyev, Aleks Shirokov, Amitav Mohanty, Daniil Balenko, Elaine Zhao, Essam Ewaisha, Hongbo Dong, Igor Marnat, Lev Novikov, Min Zeng, Steven Shingler, Timofey Durakov, Wiliam de Abreu Pinho, Ben Christensen, Mayank Pundir, Kaushik Veeraraghavan | 17m |
PIMS is a production system running five years across millions of servers and tens of thousands of services that enforces a one-domain regional contract (at… | ||
| Mimesys: Generating Realistic Executable Testing Environments from Resource Usage Traces | Donghyun Kim, Zichao Hu, Joydeep Biswas, Aditya Akella, Daehyeok Kim | 15m |
Mimesys converts time-series resource traces into executable stressor compositions represented as a per-thread 2D grid of fractional run times and pins threads… | ||
| MEGALON: Efficient Data Sharing for Partly Coherent CXL Memory | Jiyu Hu, Seokjoo Cho, Landon Johnson, Kiran Hombal, Shreesha Gopalakrishna Bhat, Marcos K. Aguilera, Ramnatthan Alagappan, Aishwarya Ganesan | 15m |
MEGALON implements a split metadata sharing design: large, rarely updated indexes are replicated in local DRAM while small, hot coherence records live in the… | ||
| Unleash All Cores: Asymmetry-Aware Scalable DNN Inference on Mobile CPUs | Qianlong Sang, Puyi He, Huanghuang Liang, Yili Gong, Chuang Hu, Xiaobo Zhou, Dazhao Cheng | 17m |
SANI implements an AMP-aware runtime atop Arm-CL that targets mobile CPUs and demonstrates concrete gains: it prepares cluster-specific kernels, dynamically… | ||
| qTPU: Hybrid Tensor Networks for Quantum-Classical Acceleration | Nathaniel Tornow, Emmanouil Giortamis, Dennis Sprokholt, Christian Mendl, Pramod Bhatotia | 16m |
qTPU introduces a hybrid tensor network abstraction plus a programming model, compiler and runtime that jointly trade quantum error against classical FLOP… | ||
| Drs.NAS: Ultra-Efficient Neural Architecture Search for Recommendation Systems | Ruixuan Wang, Xun Jiao | 17m |
Claims an automated NAS that uses a superproxy metric (seven zero-cost proxies) to avoid training and finishes search in two minutes on a commodity CPU, not… | ||
| SMARTTalk: Teaching SMART Logs to Talk to LLMs | Mayur Akewar, Dongsheng Luo, Sandeep Madireddy, Janki Bhimani | 22m |
Converts 30-day SMART windows into 5-day patches, encodes local trends with CNN patch encoders, clusters embeddings into reusable temporal patterns, and maps… | ||
| Svalinn: Overload Control in Large-Scale Servers with Multiple Resource Bottlenecks | Bhaskar Subhash Pardeshi, Peidi Song, Ahmed Saeed | 21m |
Svalinn uses a credit-based admission controller that runs online micro-experiments to find the credit pool maximizing a user-specified convex utility… | ||
| PeeR: First-Class Scheduling for Latency-Critical eBPF Applications | Jeremy Carin, Ben Holmes, Weiyang Wang, Ankit Bhardwaj, Manya Ghobadi | 16m |
PeeR makes eBPF programs preemptible and schedulable by inserting a lightweight two-cycle budget check at each helper call, saving interrupted state into a… | ||
| TypeCraft: A Lightweight Data Type Profiler with High Resolution | Zecheng Li, Xu Liu, Namhyung Kim, Blake Jones, Alexey Alexandrov, Jiajia Li | 16m |
TypeCraft maps individual memory instructions in optimized x86 C binaries to structs and fields by combining DWARF-based type resolution with a function-scoped… | ||
| Neuro-Symbolic Proof Generation for Scaling Systems Software Verification | Baoding He, Zenan Li, Wei Sun, Yuan Yao, Taolue Chen, Xiaoxing Ma, Zhendong Su | 17m |
Neuro-symbolic proof search implements a best-first tree search that repeatedly queries a fine-tuned LLM for single-step tactics, executes each step in a new… | ||
| Spain: Succinct Proofs for Numerical Computations | Zachary DeStefano, Noah Golub, Zile Huang, Julius Zhang, Sam Frank, Michael Walfish | 16m |
Spain introduces approximate constraints and a new proof protocol that represents numerical operations directly over rationals with bounded error, instead of… | ||
| RT: Regular Types for the Streaming Shell | Zekai Li, Lukas Lazarek, Evangelos Lamprou, George Kapetanakis, Konstantinos Mamouras, Nikos Vasilakis | 18m |
RT is an overlay type system that statically checks shell pipelines by modeling command input/output as regular types and reducing composition to… | ||
| VTC: DNN Compilation with Virtual Tensors for Data Movement Elimination | Muyan Hu, Ahan Gupta, Jiachen Yuan, Vima Gupta, Taeksang Kim, Xin Xu, Janardhan Kulkarni, Ofer Dekel, Vikram S. Adve, Charith Mendis | 15m |
VTC introduces virtual tensor representations that map logical indices to multiple physical buffers to eliminate intermediate data-movement kernels, implements… | ||
| nothing matches | ||