USENIX ATC '25
2025-07-07..2025-07-09
Boston, United States
as the organisers put it
The 2025 USENIX Annual Technical Conference will take place on July 7–9, 2025, at the Sheraton Boston in Boston, MA, USA. USENIX ATC brings together leading systems researchers for the presentation of cutting-edge systems research and the opportunity to gain insight into a wealth of must-know topics.
who it's for
systems researchers, machine learning engineers, infrastructure engineers, site reliability engineers, cloud platform engineers, storage engineers, network engineers
topics
large language model serving; serverless LLM platforms; smartNIC-accelerated image provisioning; multi-GPU DRL scheduling; long-context LLM fine-tuning; dynamic pipeline parallelism; GPU preemptive scheduling; disaggregated rack CXL architecture; in-network workload shaping with P4; universal distributed checkpointing; model training on Ascend chips; fail-slow hardware detection
| title | author | length |
|---|---|---|
| Joint Keynote Address: Accelerating Software Development: The LLM (R)evolution | Emery D. Berger | 44m |
A concrete recipe emerges: evolve existing tools, exploit what they uniquely know, and ensure outputs are validated. Examples: the scaline profiler++ suggests… | ||
| Fast ACS: Low-Latency File-Based Ordered Message Delivery at Scale | Sushant Kumar Gupta, Anil Raghunath Iyer, Chang Yu, Neel Bagora, Olivier Pomerleau, Vivek Kumar, Prunthaban Kanthakumar | 16m |
Fast ACS is a file-based ordered delivery system that combines file-based ordered delivery, memory caching (Clickmap L1) and Colossus L2, chunked files across… | ||
| Poby: SmartNIC-accelerated Image Provisioning for Coldstart in Clouds | Zihao Chang, Jiaqi Zhu, Haifeng Sun, Yunlong Xie, Kan Shi, Ninghui Sun, Yungang Bao, Sa Wang | 18m |
Claims SmartNIC-accelerated image provisioning reduces coldstart by offloading decompression to BlueField SmartNICs (decompress 66× faster than host) and… | ||
| Burst Computing: Quick, Sudden, Massively Parallel Processing on Serverless... | Daniel Barcelona-Pons, Aitor Arjona, Pedro López-García, Enrique Molina-Giménez, Stepan Klymonchuk | 16m |
Claims a serverless system built atop OpenWhisk that launches large jobs with a group invocation primitive (flare) to guarantee simultaneity, packs many… | ||
| DEEPSERVE: Serverless Large Language Model Serving at Scale | Junhao Hu, Jiang Xu, Zhixia Liu, Yulong He, Yuetao Chen, Hao Xu, Jiang Liu, Jie Meng, Baoquan Zhang, Shining Wan, Gengyuan Dan, Zhiyu Dong, Zhihao Ren, Changhong Liu, Tao Xie, Dayun Lin, Qin Zhang, Yue Yu, Hao Feng, Xusheng Chen, Yizhou Shan | 17m |
DEEPSERVE is a serverless LLM platform that serves models on Huawei Ascend NPUs with a concrete request-job-task model abstraction, an in-house FLOWSERVE… | ||
| Cosmic: Cost-Effective Support for Cloud-Assisted 3D Printing | Yuan Yao, Chuan He, Chinedum Okwudire, Harsha V. Madhyastha | 16m |
Cloud controller called Cosmic partitions serverless workers into groups and issues timely speculation to hide ~10 ms invocation delays measured on AWS Lambda… | ||
| GMI-DRL: Empowering Multi-GPU DRL with Adaptive-Grained Parallelism | Yuke Wang, Boyuan Feng, Zheng Wang, Guyue Huang, Ang Li, Yufei Ding | 16m |
GMI-DRL introduces a sub-GPU abstraction called GPU Multiplexing Instance (GMI) plus an adaptive Coordinator and a specialized Communicator to map… | ||
| mTuner: Accelerating Parameter-Efficient Fine-Tuning on Multi-GPU Servers with... | Kezhao Huang, Siqi Zhu, Mingshu Zhai, Liyan Zheng, Kinman Lei, Jiaao He, Yuyang Jin, Jidong Zhai | 16m |
mTuner implements an Elastic Tensor abstraction that exposes four tunable operations — gather, discard, execute, checkpoint — and a one‑time schedule search to… | ||
| JENGA: Enhancing LLM Long-Context Fine-tuning with Contextual Token Sparsity | Tuowei Wang, Xingyu Chen, Kun Li, Ting Cao, Ju Ren, Yaoxue Zhang | 15m |
JENGA identifies and prunes redundant tokens during long-context fine-tuning to cut activation memory and speed up training, claiming up to 1.93× memory… | ||
| FlexPipe: Maximizing Training Efficiency for Transformer-based Models with... | Hairui Zhao, Qi Tian, Hongliang Li, Zizhong Chen | 14m |
FlexPipe is a middleware that dynamically resizes pipeline parallelism per iteration, adding or removing PP stages without stalling via a live flexibility… | ||
| Optimus: Accelerating Large-Scale Multi-Modal LLM Training by Bubble Exploitation | Weiqi Feng, Yangrui Chen, Shaoyu Wang, Yanghua Peng, Haibin Lin, Minlan Yu | 15m |
Reports a system called Optimus that cuts end-to-end MLLM training time by ~20.5–21.3% on 3,072 GPUs (ViT-22B encoder + GPT-175B backbone) by turning idle GPU… | ||
| Towards Optimal Rack-scale μs-level CPU Scheduling through In-Network Workload... | Xudong Liao, Han Tian, Xinchen Wan, Chaoliang Zeng, Hao Wang, Junxue Zhang, Mengyu Ma, Guyue Liu, Kai Chen | 16m |
Pallas partitions rack traffic at the ToR so each shard has homogeneous CPU demand, uses a P4-programmed ToR switch for classification and per-group JSQ-style… | ||
| TGW: Operating an Efficient and Resilient Cloud Gateway at Scale | Yifan Yang, Lin He, Jiasheng Zhou, Xiaoyi Shi, Yichi Xu, Shicheng Wang, Jinlong E, Ying Liu, Junwei Zhang, Zhuang Yuan, Hengyang Xu | 18m |
DPDK-based forwarding plane achieves 2.9× single-node forwarding over prior systems; deployment uses load distributors (LDs) with per-flow core pinning for EIP… | ||
| MARC: Motion-Aware Rate Control for Mobile E-commerce Cloud Rendering | Yuankang Zhao, Furong Yang, Gerui Lv, Qinghua Wu, Yanmei Liu, Jiuhai Zhang, Yutang Peng, Feng Peng, Hongyu Guo, Ying Chen, Zhenyu Li, Gaogang Xie | 15m |
MARC is a server-side, frame-level rate controller that predicts short-term motion with a lightweight Markov model and uses frame-level stochastic optimization… | ||
| Accelerating Distributed Graph Learning by Using Collaborative In-Network... | Zhaoyi Li, Jiawei Huang, Yijun Li, Jingling Liu, Junxue Zhang, Hui Li, Xiaojun Zhu, Shengwen Zhou, Jing Shao, Xiaojuan Lu, Qichen Su, Jianxin Wang, Chee Wei Tan, Yong Cui, Kai Chen | 16m |
Offloads GNN multicast and many‑to‑one aggregation into the network using a P4 programmable switch and DPDK host stack, implements graph-aware multicast… | ||
| Opening Up Kernel-Bypass TCP Stacks | Shinichi Awamoto, Michio Honda | 17m |
Six kernel and kernel-bypass TCP stacks are compared on identical hardware and the same HTTP workload with per-stack optimized servers, disabled NIC offloads… | ||
| GPREEMPT: GPU Preemptive Scheduling Made General and Efficient | Ruwen Fan, Tingxu Ren, Minhui Xie, Shiwei Gao, Jiwu Shu, Youyou Lu | 12m |
Repurposes GPU hardware time slicing and a small driver API to achieve near-instant context-switch preemption, claiming preemption latency within 40 μs while… | ||
| μEFI: A Microkernel-Style UEFI with Isolation and Transparency | Le Chen, Yiyang Wu, Jinyu Gu, Yubin Xia, Haibo Chen | 15m |
A microkernel-style UEFI isolates drivers by deprivileging modules to user mode and placing each in its own address space, implements trampoline injection and… | ||
| PageFlex: Flexible and Efficient User-space Delegation of Linux Paging Policies... | Anil Yelam, Kan Wu, Zhiyuan Guo, Suli Yang, Rajath Shashidhara, Wei Xu, Stanko Novaković, Alex C. Snoeren, Kimberly Keeton | 12m |
PageFlex externalizes Linux paging policies by running eBPF policy handlers that read low-overhead in-kernel signals, store a small per-page state slot, and… | ||
| ASTERINAS: A Linux ABI-Compatible, Rust-Based Framekernel OS with a Small and... | Yuke Peng, Hongliang Tian, Junyang Zhang, Ruihan Li, Chengjun Chen, Jianfeng Jiang, Jinyi Xian, Xiaolin Wang, Chenren Xu, Diyu Zhou, Yingwei Luo, Shoumeng Yan, Yinqian Zhang | 18m |
ASTERINAS is a Linux ABI–compatible OS implemented in safe Rust on a novel framekernel architecture with an OSTD crate that enforces intra-kernel privilege… | ||
| Rex: Closing the language-verifier gap with safe and usable kernel extensions | Jinghao Jia, Ruowen Qin, Milo Craun, Egor Lukiyanov, Ayush Bansal, Minh Phan, Michael V. Le, Hubertus Franke, Hani Jamjoom, Tianyin Xu, Dan Williams | 19m |
Rex claims to close the language‑verifier gap by running kernel extensions in safe Rust linked to a Rex kernel crate and a lightweight runtime that enforces… | ||
| Barre: Empowering Simplified and Versatile Programmable Congestion Control in... | Yajuan Peng, Haoran Wei, Xiaolong Zhong, Junkai Huang, Haohan Xu, Zicheng Wang, Yang Bai, Zhuo Jiang, Jianxi Ye, Xiaoliang Wang, Xiaoming Fu, Huichen Dai | 17m |
Barre is a practical PCC-style congestion controller that runs on BlueField-3 SmartNICs, uses RTD probe packets plus high-frequency CMP signals, and implements… | ||
| FLB: Fine-grained Load Balancing for Lossless Datacenter Networks | Jinbin Hu, Wenxue Li, Xiangzhou Liu, Junfeng Wang, Bowen Liu, Ping Yin, Jianxin Wang, Jiawei Huang, Kai Chen | 16m |
FLB claims to balance RoCE traffic by combining threshold-free rerouting for uncongested flows with congested-flow isolation when queues near PFC thresholds… | ||
| SNARY: A High-Performance and Generic SmartNIC-accelerated Retrieval System | Qiaoyin Gan, Heng Pan, Luyang Li, Kai Lv, Hongtao Guan, Zhaohua Wang, Zhenyu Li, Gaogang Xie | 13m |
SNARY implements a SmartNIC-accelerated retrieval engine using on-card HBM and a data-parallel exact search plus an LSH-based fuzzy search, deployed on Xilinx… | ||
| Minos : A Lightweight and Dynamic Defense against Traffic Analysis in... | Zihao Wang, Qing Li, Guorui Xie, Dan Zhao, Kejun Li, Zhuochen Fan, Lianbo Ma, Yong Jiang | 18m |
Minos implements a programmable-switch defense on Tofino1 that performs line‑rate header encryption and flow mixing to defend website and IoT fingerprinting… | ||
| GeneralSparse: Bridging the Gap in SpMM for Pruned Large Language Model Inference.. | Yaoyu Wang, Xiao Guo, Junmin Xiao, De Chen, Guangming Tan | 19m |
GeneralSparse is an automated SPMM generator for GPUs that maps diverse pruning patterns (random and magnitude) to an orthogonal memory-division plus reduction… | ||
| HyCache: Hybrid Caching for Accelerating DNN Input Preprocessing Pipelines | Keshav Vinayak Jha, Shweta Pandey, Murali Annavaram, Arkaprava Basu | 14m |
Hybrid Cache automatically profiles ~1% of a dataset, models fetch/decode/augment costs, and uses an ILP solver to pick per-stage, tier-aware placements that… | ||
| The Koala Benchmarks for the Shell: Characterization and Implications | Evangelos Lamprou, Ethan Williams, Georgios Kaoukis, Zhuoxuan Zhang, Michael Greenberg, Konstantinos Kallas, Lukas Lazarek, Nikos Vasilakis | 13m |
Koala is a curated benchmark suite of real POSIX shell programs with permanently stored inputs and automated, configurable tooling that installs dependencies… | ||
| KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a... | Jiahao Wang, Jinbo Han, Xingda Wei, Sijie Shen, Dingyan Zhang, Chenguang Fang, Rong Chen, Wenyuan Yu, Haibo Chen | 16m |
Paper analyzes a one-week Alibaba trace with per-request timestamps, modalities, turns and KV hashes and finds reuse is skewed: single-turn requests contribute… | ||
| LogCrisp: Fast Aggregated Analysis on Large-scale Compressed Logs by Enabling... | Junyu Wei, Guangyan Zhang, Junchao Chen, Qi Zhou | 09m |
LogCrisp implements a two-phase pattern extraction that extracts a global Sketch offline and lightweight per-block Specs online, decoupling global description… | ||
| HotRAP: Hot Record Retention and Promotion for LSM-trees with Tiered Storage | Jiansheng Qiu, Fangzhou Yuan, Mingyu Gao, Huanchen Zhang | 12m |
HotRAP is a RocksDB-based key-value store that implements record-level promotion using an additional small on-disk LSM called route to track hotness with… | ||
| Mitigating Resource Usage Dependency in Sorting-based KV Stores on Hybrid... | Qingyang Zhang, Yongkun Li, Yubiao Pan, Haoting Tang, Yinlong Xu | 15m |
DecouKV decouples sorting into index merge tasks and data append and data flush tasks by separating indexes (stored in a mergeable skip list) from values… | ||
| SolFS: An Operation-Log Versioning File System for Hash-free Efficient Mobile... | Riwei Pan, Yu Liang, Lei Li, Hongchao Du, Tei-Wei Kuo, Chun Jason Xue | 20m |
SolFS implements an operation-log versioning file system that records each write as per-file mergeable operation logs (offset+length stored in an extent tree)… | ||
| Z-LFS: A Zoned Namespace-tailored Log-structured File System for Commodity Small... | Inhwi Hwang, Sangjin Lee, Sunggon Kim, Hyeonsang Eom, Yongseok Son | 16m |
Z-LFS is a Linux-kernel file system built on F2FS for commodity small-zone ZNS SSDs that implements append-only metadata, speculative log stream management… | ||
| CLONE: Customizing LLMs for Efficient Latency-Aware Inference at the Edge | Chunlin Tian, Xinpeng Qin, Kahou Tam, Li Li, Zijian Wang, Yuanzhe Zhao, Minglei Zhang, Chengzhong Xu | 17m |
Claims a combined algorithm–hardware co‑design that prunes and customizes LLMs to fit edge memory/energy budgets, implements a 28nm scalable accelerator with a… | ||
| Weaver: Efficient Multi-LLM Serving with Attention Offloading | Shiwei Gao, Qing Wang, Shaoxun Zeng, Youyou Lu, Jiwu Shu | 10m |
Weaver implements attention offloading that lets hot models borrow KV cache memory on cold-model GPUs by shipping QKV vectors and running attention there… | ||
| Torpor: GPU-Enabled Serverless Computing for Low-Latency, Resource-Efficient... | Minchen Yu, Ao Wang, Dong Chen, Haoxuan Yu, Xiaonan Luo, Zhuohao Li, Wei Wang, Ruichuan Chen, Dapeng Nie, Haoran Yang, Yu Ding | 16m |
Torpor implements late binding with model swapping to keep models in host memory and move them to a shared GPU pool on demand, using asynchronous API… | ||
| Toppings: CPU-Assisted, Rank-Aware Adapter Serving for LLM Inference | Suyi Li, Hanfeng Lu, Tianyuan Wu, Minchen Yu, Qizhen Weng, Xusheng Chen, Yizhou Shan, Binhang Yuan, Wei Wang | 14m |
System pins the base LLM on GPUs and dynamically loads LoRA adapters from host memory, while CPU-assisted LoRA computation precomputes the lightweight XAB… | ||
| QFactory: Accelerating Quantized Large Language Model Serving with Qtile Graphs | Qihao Zhang, Mingshu Zhai, Rui Sun, Jidong Zhai | 14m |
QFactory is a compilation framework that generates fused quantized kernels by deferring dequantization, representing quantized tensors with a Qtile-graph… | ||
| PluS: Highly Efficient and Expandable ML Compiler with Pluggable Graph Schedules | Ruofan Wu, Zhen Zheng, Feng Zhang, Chuanjie Liu, Zaifeng Pan, Jidong Zhai, Xiaoyong Du | 15m |
PluS implements a loop-centric plus graph abstraction plus a pattern warehouse for decoupled subgraph matching and a PluS code template interface, letting… | ||
| Obscura: Concealing Recomputation Overhead in Training of Large Language Models... | Yuzhou Huang, Yapeng Jiang, Zicong Hong, Wuhui Chen, Bin Wang, Weixi Zhu, Yue Yu, Zibin Zheng | 20m |
Obscura is a pipeline training system that transforms the standard 1F1B schedule to shift forward work into unused forward bubbles and turn them into backward… | ||
| PPipe: Efficient Video Analytics Serving on Heterogeneous GPU Clusters via... | Z. Jonny Kong, Qiang Xu, Y. Charlie Hu | 17m |
PPipe implements pool-based pipeline parallelism to split CNN models across heterogeneous GPUs, uses an MILP-based control plane to pick partition points and… | ||
| Voltrix: Sparse Matrix-Matrix Multiplication on Tensor Cores with Asynchronous... | Yaqi Xia, Weihu Wang, Donglin Yang, Xiaobo Zhou, Dazhao Cheng | 19m |
Voltrix-SpMM implements a CUDA 12.6 kernel for H100 that combines a bitwise-compressed BMAC format (128-bit blocks loaded with LDGXT) with TMA-driven… | ||
| NetKeeper: Enhancing Network Resilience with Autonomous Network Configuration... | Zhaoyang Wan, Rongxin Han, Haifeng Sun, Qi Qi, Zirui Zhuang, Bo He, Liang Zhang, Jianxin Liao, Jingyu Wang | 18m |
NetKeeper implements northbound and southbound interfaces plus a DSL and interpreter that accept natural language and anomaly logs parsed by a fine‑tuned… | ||
| GREYHOUND: Hunting Fail-Slows in Hybrid-Parallel Training at Scale | Tianyuan Wu, Wei Wang, Yinghao Yu, Siran Yang, Wenchao Wu, Qinkai Duan, Guodong Yang, Jiamang Wang, Lin Qu, Liping Zhang | 18m |
Characterization on a 10,000+ GPU multi-tenant cluster finds transient fail-slows from CPU contention, GPU thermal throttling, and network congestion that last… | ||
| Crash Consistency in Block-Level Caching Systems: An Open CAS Case Study | Shaohua Duan, Youmin Chen | 18m |
Paper shows Open CAS fails end-to-end crash guarantees: a workload-oriented testing harness injects synthetic crashes at precise cache states (including… | ||
| FiDe: Reliable and Fast Crash Failure Detection to Boost Datacenter Coordination | Davide Rovelli, Pavel Chuprikov, Philipp Berdesinski, Ali Pahlevan, Patrick Jahnke, Patrick Eugster | 15m |
FiDe reports crashes in datacenters in under 30 μs (7.2× faster than prior art) by creating stable end-to-end interactions: FiDe reports crashes <30 μs (7.2×… | ||
| LEOCraft: Towards Designing Performant LEO Networks | Suvam Basak, Amitangshu Pal, Debopam Bhattacherjee | 16m |
LEOCraft is a modular, flow‑level LEO network simulator and visualization that computes throughput, coverage, latency and hop counts and renders topology… | ||
| Emulating Space Computing Networks with RHONE | Liying Wang, Qing Li, Yuhan Zhou, Zhaofeng Luo, Donghao Zhang, Shangguang Wang, Xuanzhe Liu, Chenren Xu | 16m |
RHONE emulates Space Computing Networks by building offline models of power, thermal, orbit, network, and computation from real telemetry and… | ||
| Roaming Free in the VR World with MP2 | Yifei Xu, Xumiao Zhang, Yuning Chen, Pan Hu, Xuan Zeng, Zhilong Zheng, Xianshang Lin, Yanmei Liu, Songwu Lu, Z. Morley Mao, Wan Du, Dennis Cai, Ennan Zhai, Yunfei Ma | 19m |
A user-space centralized overlay called MP2 coordinates per-headset two Wi‑Fi interfaces via a centralized controller that collects RSSI and frame-size stats… | ||
| STORM: a Multipath QUIC Scheduler for Quick Streaming Media Transport under... | Liekun Hu, Changlong Li | 20m |
A multipath QUIC scheduler called STORM runs on mobile devices (evaluated on a Xiaomi T12 Pro using WiFi and 5G) that differentiates reliable versus unreliable… | ||
| Internet Connection Splitting: What’s Old is New Again | Gina Yuan, Thea Rossman, Keith Winstein | 14m |
Emulation study uses TCNETM and a novel split-throughput heuristic to estimate PEP gains across 150 emulations covering an estimated 8,000 split settings. Key… | ||
| WIC: Hiding Producer-Consumer Synchronization Delays with Warp-Level... | Jiajian Zhang, Fangyu Wu, Hai Jiang, Qiufeng Wang, Genlang Chen, Chaoyi Pang | 20m |
WIC replaces consumer-side polling with warp-level interrupts that preempt spinning warps and park them in the UVM host kernel, using UVM page fault monitoring… | ||
| Primus: Unified Training System for Large-Scale Deep Learning Recommendation Models | Jixi Shan, Xiuqi Huang, Yang Guo, Hongyue Mao, Ho-Pang Hsu, Hang Cheng, Can Wang, Jun Song, Rui Shi, Xiaofeng Gao, Jingwei Xu, Shiru Ren, Jiaxiao Zheng, Hua Huang, Lele Yu, Peng Xu, Guihai Chen | 17m |
Primus implements a unified resource controller that standardizes YARN and Kubernetes semantics, runs dynamic horizontal+vertical scaling across multi‑cluster… | ||
| Chitu: Avoiding Unnecessary Fallback in Byzantine Consensus | Rongji Huang, Xiangzhe Wang, Xiaofeng Yan, Lei Fan, Guangtao Xue, Shengyun Liu | 18m |
Chitu is an asynchronous DAG BFT protocol that tries a fast path based only on message exchange and falls back to coin or leader only when proposals diverge… | ||
| Fast Distributed Transactions for RDMA-based Disaggregated Memory | Haodi Lu, Haikun Liu, Yujian Zhang, Zhuohui Duan, Xiaofei Liao, Hai Jin, Yu Zhang | 19m |
HDTX is a distributed transaction system for RDMA disaggregated memory that claims up to 88.3% and 72.1% lower latency and up to 2.08× and 84.7% higher… | ||
| Cuckoo for Clients: Disaggregated Cuckoo Hashing | Stewart Grant, Alex C. Snoeren | 18m |
RCuckoo is a fully disaggregated, lock‑based key/value store that uses one‑sided RDMA to get single round‑trip reads for small values and two round trips for… | ||
| LITESHIELD: Secure Containers via Lightweight, Composable Userspace μKernel... | Kaesi Manakkal, Nathan Daughety, Marcus Pendleton, Hui Lu | 18m |
LITESHIELD claims VM-like isolation while avoiding VMs by reducing the user‑to‑host interface to 22 syscalls. It implements fast shared‑memory IPC and syscall… | ||
| Accelerating Nested Virtualization with HyperTurtle | Ori Ben Zur, Jakob Krebs, Shai Aviram Bergman, Mark Silberstein | 16m |
HyperTurtle reduces nested-VM world switches by offloading L1 code as eBPF running in L0, keeping L1 control while making only a few hundred non-intrusive… | ||
| Efficient Performance-Aware GPU Sharing with Compatibility and Isolation through... | Shulai Zhang, Ao Xu, Quan Chen, Han Zhao, Weihao Cui, Zhen Wang, Yan Li, Limin Xiao, Minyi Guo | 14m |
KRYPTON implements kernel-space command-buffer interception and a spatio-temporal orchestrator on A100 and 4090 hardware, using offline profiling, CPU tokens… | ||
| AnchorNet: Bridging Live and Collaborative Streaming with a Unified Architecture | Tong Meng, Wei Zhang, Dong Chen, Zhen Wang, Quanqing Li, Changqing Yan, Wei Yang, Chao Yuan, Le Zhang, Jianxin Kuang, Jianlin Xu | 14m |
AnchorNet implements a unified stream path with the RTC SFU as the single CDN publisher and server-side mixing, solving mode switches between single-host and… | ||
| Efficient Workflow Serving for Diffusion Models with Many Adapters | Katz | 12m |
System decouples heavy ControlNets from the base diffusion model by deploying them on dedicated GPUs as ControlNet-as-a-Service, enabling caching, parallel… | ||
| PopFetcher: Towards Accelerated Mixture-of-Experts Training Via Popularity... | Junyi Zhang, Chuanhu Ma, Xiong Wang, Yuntao Nie, Yuqing Li, Yuedong Xu, Xiaofei Liao, Bo Li, Hai Jin | 20m |
PopFetcher implements sliding-window popularity prediction to identify hot experts and does expert-wise prefetching of their parameters into local GPU memory… | ||
| HypeReca: Distributed Heterogeneous In-Memory Embedding Database for Training... | Jiaao He, Shengqi Chen, Kezhao Huang, Jidong Zhai | 22m |
HypeReca is a distributed in-memory key‑value embedding database that uses both GPU and CPU memory, implements decentralized indexing tables (DIT) with a local… | ||
| CrossPipe: Towards Optimal Pipeline Schedules for Cross-Datacenter Training | Tiancheng Chen, Ales Kubicek, Langwen Huang, Torsten Hoefler | 19m |
CrossPipe explicitly models cross-datacenter latency and bandwidth with an alpha–beta cost model and produces solver-based optimal or greedy near-optimal… | ||
| Unveiling Compiler Faults via Attribute-Guided Compilation Space Exploration | Jiangchang Wu, Yibiao Yang, Maolin Sun, Yuming Zhou | 14m |
ATLAS inserts attributes into functions, variables, and statements, analyzes types and local-variable relations to pick compatible compiler flags, and compiles… | ||
| Understanding and Detecting Fail-Slow Hardware Failure Bugs in Cloud Systems | Gen Dong, Yu Hua, Yongle Zhang, Zhangyu Chen, Menglei Chen | 14m |
Claims a systematic study of 48 real-world failures and finds fine-grained, fail-slow hardware behavior repeatedly breaks synchronized and timeout-protected… | ||
| Converos: Practical Model Checking for Verifying Rust OS Kernel Concurrency | Ruize Tang, Minghua Wang, Xudong Sun, Lin Huang, Yu Huang, Xiaoxing Ma | 15m |
Claims a practical workflow that verifies 12 Rust OS concurrency modules (≈4k LOC) using PlusCal/TLC and lightweight logging, finds 20 real bugs (9 via model… | ||
| Bin2Wrong: a Unified Fuzzing Framework for Uncovering Semantic Errors in Binary... | Zao Yang, Stefan Nagy | 15m |
BIN2WRONG implements a decompiler-agnostic fuzzer that mutates source, compiler options, optimizations and binary format via a unified testcase structure… | ||
| HEC: Equivalence Verification Checking for Code Transformation via Equality... | Jiaqi Yin, Zhan Song, Nicolas Bohm Agostini, Antonino Tumeo, Cunxi Yu | 15m |
HEC introduces a concrete equivalence checker that leverages an e-graph backend on MLIR, combining over 60 static rewrite rules with generated dynamic rewrite… | ||
| Para-ksm: Parallelized Memory Deduplication with Data Streaming Accelerator | Houxiang Ji, Minho Kim, Seonmu Oh, Daehoon Kim, Nam Sung Kim | 17m |
Para-ksm implements a redesigned KSM that uses Intel Data Streaming Accelerator to offload page comparisons and checksum computation and batches multiple pages… | ||
| DSA-2LM: A CPU-Free Tiered Memory Architecture with Intel DSA | Ruili Liu, Teng Ma, Mingxing Zhang, Jialiang Huang, Yingdi Shan, Zheng Liu, Lingfeng Xiang, Zhen Lin, Hui Lu, Jia Rao, Kang Chen, Yongwei Wu | 17m |
DSA-2LM integrates Intel DSA into Linux kernel page migration, bypasses VA translation to submit physical addresses, and implements a concurrent migration… | ||
| Turbocharge ANNS on Real Processing-in-Memory by Enabling Fine-Grained Per-PIM... | Puqing Wu, Minhui Xie, Enrui Zhao, Dafang Zhang, Jing Wang, Xiao Liang, Kai Ren, Yunpeng Chai | 19m |
PIMANN retrofits UPMEM by tapping an undocumented per-PU control interface to switch DDR ownership and run a persistent PIM kernel that eliminates inter-batch… | ||
| SwCC: Software-Programmable and Per-Packet Congestion Control in RDMA Engine | Hongjing Huang, Jie Zhang, Xuzheng Chen, Ziyu Song, Jiajun Qin, Zeke Wang | 19m |
SwCC embeds on-NIC RISC-V cores and a QP-aware memory subsystem on a Xilinx U280 FPGA to run per-packet congestion control in software while matching hardware… | ||
| DRack: A CXL-Disaggregated Rack Architecture to Boost Inter-Rack Communication | Xu Zhang, Ke Liu, Yuan Hui, Xiaolong Zheng, Yisong Chang, Yizhou Shan, Guanghui Zhang, Ke Zhang, Yungang Bao, Mingyu Chen, Chenxi Wang | 17m |
Proposes disaggregating all NICs and memory inside a rack into a shared NIC pool and memory pool accessed with CXL semantics, so any host can drive multiple… | ||
| ShieldReduce: Fine-Grained Shielded Data Reduction | Jingyuan Yang, Jun Wu, Ruilin Wu, Jingwei Li, Patrick P. C. Lee, Xiong Li, Xiaosong Zhang | 17m |
System design applies Intel SGX enclave to perform deduplication, delta and local compression inside a trusted environment, using a novel bi-directional delta… | ||
| MemoryTrap: Booby Trapping Memory to Counter Memory Disclosure Attacks with... | Chenke Luo, Jiang Ming, Dongpeng Xu, Guojun Peng, Jianming Fu | 15m |
MemoryTrap compiles applications to sprinkle unreadable booby traps into code and uses Intel MPK plus a kernel exception handler to detect and stop memory… | ||
| Separate but Together: Integrating Remote Attestation into TLS | Carsten Weinhold, Muhammad Usama Sardar, Ionut Mihalcea, Yogesh Deshpande, Hannes Tschofenig, Yaron Sheffer, Thomas Fossati, Michael Roitzsch | 16m |
Protocol integrates TLS 1.3 extensions with remote attestation by hashing handshake messages into a new linking hash that the TEE signs alongside the TLS… | ||
| DDLumos: Understanding and Detecting Atomic DDL Bugs in DBMSs | Zhiyong Wu, Jie Liang, Jingzhou Fu, Wenqian Deng, Yu Jiang | 15m |
Analyzes 207 real-world Atomic DDL bugs and builds DDLUMOS, a tester that uses metadata conflict-guided DDL synthesis and graph-based consistency analysis to… | ||
| SpaceExit: Enabling Efficient Adaptive Computing in Space with Early Exits | Jiacheng Liu, Xiaozhi Zhu, Tongqiao Xu, Xiaofeng Hou, Chao Li | ? |
System builds an onboard pipeline that tiles images, runs a multiscale feature backbone with early exits, and schedules tiles across heterogeneous accelerators… | ||
| XRT: An Accelerator-Aware Runtime for Accelerated Chip Multiprocessors | Neel Patel, Mohammad Alian | 15m |
XRT is a runtime that runs on accelerated CMPs and claims up to 3.2× higher throughput-under-SLO versus an unoptimized runtime by distributing accelerator… | ||
| DShuffle: DPU-Optimized Shuffle Framework for Large-scale Data Processing | Chen Ding, Sicen Li, Kai Lu, Ting Yao, Daohui Wang, Huatao Wu, Jiguang Wan, Zhihu Tan, ChangSheng Xie | 19m |
DShuffle fully offloads Spark shuffle onto Bloomfield 3 DPUs using DPA-accelerated serialization and DPU direct spilling via PCIe P2P, and organizes work into… | ||
| Accelerating Model Training on Ascend Chips: An Industrial System for Profiling... | Yuhang Zhou, Zibo Wang, Zhibin Wang, Ruyi Zhang, Chen Tian, Xiaoliang Wang, Wanchun Dou, Guihai Chen, Bingqiang Wang, Yonghong Tian, Yan Zhang, Hui Wang, Fuchun Wei, Boquan Sun, Jingyi Zhang, Bin She, Teng Su, Yifan Yao, Chunsheng Li, Ziyang Zhang, Yaoyuan Wang, Bin Zhou, Guyue Liu | 13m |
Hermes is an industrial system that combines cost-to-find profiling (lightweight monitor that triggers fine-grained traces) with a hierarchical inter- and… | ||
| CAFault: Enhance Fault Injection Technique in Practical Distributed Systems via... | Yuanliang Chen, Fuchen Ma, Yuanhang Zhou, Zhen Yan, Yu Jiang | 19m |
A framework that generates and prunes FDModel configurations by observing coverage changes and then drives focused testing with fault-handling guided fuzzing… | ||
| Revealing Floating-Point Accumulation Orders in Software/Hardware Implementations | Peichen Xie, Yanjie Gao, Yang Wang, Jilong Xue | 19m |
A diagnostic tool called FPRev nonintrusively infers the actual floating point accumulation tree by feeding specially crafted inputs that exploit precision… | ||
| Inferring Likely Counting-related Atomicity Program Properties for Persistent... | Yunmo Zhang, Junqiao Qiu, Hong Xu, Chun Jason Xue | 17m |
Method infers counting-related atomicity properties for persistent memory by encoding an access-range invariance between container indices and size counters… | ||
| Optimizing Input Minimization in Kernel Fuzzing | Hui Guo, Hao Sun, Shan Huang, Ting Su, Geguang Pu, Shaohua Li | 18m |
Introduces SyzMini, a Syzkaller patch that implements influence-guided call removal and type-informed argument simplification to cut minimization executions… | ||
| IRHash: Efficient Multi-Language Compiler Caching by IR-Level Hashing | Tobias Landsberg, Johannes Grunenberg, Christian Dietrich, Daniel Lohmann | 16m |
IRHash implements IR-level hashing as an LLVM pass that stores/retrieves object files by hashing the generated IR, supporting any language with an LLVM… | ||
| On-Demand Container Partitioning for Distributed ML | Giovanni Bartolomeo, Navidreza Asadi, Wolfgang Kellerer, Jorg Ott, Nitinder Mohan | 19m |
A complete build-and-distribution stack that adds a new layer type for big files and a 2DFS two-dimensional filesystem enabling independent allotments, a… | ||
| PathWeaver: A High-Throughput Multi-GPU System for Graph-Based Approximate... | Sukjin Kim, Seongyeon Park, Si Ung Noh, Junguk Hong, Taehee Kwon, Hunseong Lim, Jinho Lee | 16m |
PathWeaver is a multi-GPU ANNS system that implements pipeline-based path extension and ghost staging to pass local search results between GPUs and a tiny… | ||
| Universal Checkpointing: A Flexible and Efficient Distributed Checkpointing... | Xinyu Lian, Sam Ade Jacobs, Lev Kurilenko, Masahiro Tanaka, Stas Bekman, Olatunji Ruwase, Minjia Zhang | 15m |
Universal Checkpointing decouples on-disk state from model parallel layout by converting checkpoints into an atomic checkpoint format that stores parameters… | ||
| Towards High-Performance Transactional Stateful Serverless Workflows with... | Jianjun Zhao, Haikun Liu, Shuhao Zhang, Haodi Lu, Yancan Mao, Zhuohui Duan, Xiaofei Liao, Hai Jin | 17m |
RTSFaaS is an RDMA-capable FaaS framework that implements a lease-based concurrency protocol, builds a global task-precedence graph, and transfers leases with… | ||
| Swift: Fast Performance Tuning with GAN-Generated Configurations | Chao Chen, Shixin Huang, Xuehai Qian, Zhibin Yu | 18m |
Swift combines GAN-generated configurations with random sampling inside a Bayesian Optimization loop to skew the search toward high-performing settings… | ||
| PMR: Fast Application Response via Parallel Memory Reclaim on Mobile Devices | Wentong Li, Li-Pin Chang, Yu Mao, Liang Shi | 14m |
PMR is a kernel patch that decouples page shrinking from writeback and implements proactive page shrinking (PPS) plus storage-friendly page writeback (SPW) to… | ||
| SAVE: Software-Implemented Fault Tolerance for Model Inference against GPU Memory.. | Wenxin Zheng, Bin Xu, Jinyu Gu, Haibo Chen | 15m |
Claims software-only SAVE keeps model accuracy even under 4,000 GPU memory bit flips while adding less than 9% runtime overhead and under 1% correctness… | ||
| Identifying and Analyzing Pitfalls in GNN Systems | Yidong Gong, Arnab Kanti Tarafder, Saima Afrin, Pradeep Kumar | 15m |
Analysis finds that omitting training accuracy hides system bugs and yields misleading optimizations: framework runtime overhead dominates for small graphs… | ||
| Bluetooth Low Energy Security Testing with Combinatorial Methods | Dominik-Philip Schreiber, Manuel Leithner, Jovan Zivanovic, Dimitris E. Simos | 16m |
Implements combinatorial security testing on BLE by generating per-layer covering arrays (t=2) and combining them into a meta-CA, driving a modified Greyhound… | ||
| Resource Multiplexing in Tuning and Serving Large Language Models | Yongjun He, Haofeng Yang, Yao Lu, Ana Klimovic, Gustavo Alonso | 15m |
LLMStation is a runtime that co-schedules parameter-efficient fine-tuning and inference on the same GPU by combining iteration-level multitasking scheduling, a… | ||
| Colocating ML Inference and Training with Fast GPU Memory Handover | Jiali Wang, Yankui Wang, Mingcong Han, Rong Chen | 15m |
A new system called SIRIUS colocates latency-sensitive inference and throughput-heavy training on the same GPU by dynamically resizing training mini-batches… | ||
| AssyLLM: Efficient Federated Fine-tuning of LLMs via Assembling Pre-trained Blocks | Shichen Zhan, Li Li, Chengzhong Xu | 21m |
AssyLLM assembles pre-trained transformer blocks from multiple LLMs (LLaMA, OPT, etc.) so clients only run forward passes and upload block indices instead of… | ||
| Learning-Enhanced High-Throughput Pattern Matching Based on Programmable Data Plane | Guanglin Duan, Yucheng Huang, Zhengxin Zhang, Qing Li, Dan Zhao, Zili Meng, Dirk Kutscher, Ruoyu Li, Yong Jiang, Mingwei Xu | 18m |
Trochilus converts existing rule sets into a byte-level recurrent neural network BRNN, distills that model into a hardware-friendly soft multi-view forest… | ||
| nothing matches | ||