Adaptive quantization cuts mobile LLM cold starts by 4x
EdgeFlow lowers precision on less critical weights to speed loading from flash while preserving accuracy on phone NPUs.
cs.OS
Operating Systems
Roughly includes material in ACM Subject Classes D.4.1, D.4.2., D.4.3, D.4.4, D.4.5, D.4.7, and D.4.9.
sort pith recommended most recent
Time-to-live KV caches cut agent latency 3.7x
Keeping a finished turn's KV cache alive across short tool calls removes prefill and per-turn queueing delays.
· “Continuum: Efficient and Robust Multi-Turn LLM Agent Scheduling with KV Cache Time-to-Live”
Trusted apps can reprogram the SSD's FTL at 3.74% overhead
A new meta-interface lets trusted administrators install on-device storage policies even when the OS is hostile.
· “SxSSD: A Secure and Extensible Software-defined Solid State Drive”
mold links 10 GiB debug binaries in seconds
By running every pass in parallel, mold cuts the link step for large C++ builds from minutes to seconds.
Single weak edge GPU: 17 concurrent V-MoE adaptations vs
Trading isolation for co-adaptation cuts latency up to 58% and lifts SLO attainment 1.6x, throughput 2.1x.
· “AdaSprite: Resource-efficient Online Co-Adaptation for V2I Systems Under Large-scale Data Drifts”
Tools and state, not just models, set agent serving costs
Across ten apps and three production traces, non-LLM work dominates latency, memory, and cost; coordinated serving cuts latency 29-40%.
· “From LLM Inference to Agentic Workloads: Characterization and Implications for Serving Systems”
Padding restores deterministic LLM output at 15-324% higher throughput
CoRun isolates prefill and pads decode, so the same prompt yields the same tokens no matter what runs alongside it.
· “CoRun: Padding is Simple and Efficient for Deterministic LLM Inference”
One-shot memory reclaim cuts compute p99 by 6%
A PSI-triggered 96 MiB write to cgroup v2 improves compute tail latency but leaves SQLite unchanged, so gains are workload-specific.
· “A Bounded Reclaim Actuator for PSI-Guided Compressed Memory: A Controlled Ablation”
vToken reclaims partially filled KV blocks
Token-level eviction now reclaims up to 72.3% of retained KV blocks per request.
· “vToken: Token-Level Virtualization for Reclaimable KV Caches”
A core handoff between VMs in 13 microseconds
HyperFlux makes VM core counts elastic at runtime, protecting bursty tenants' tail latency without overprovisioning.
· “Offering Microsecond-Scale Cross-VM Core Elasticity on Colocated Lightweight Virtual Machines”
Exact packing finds 12.8% more GPU-ready agent work than fixed windows
A frozen 851-session replay shows fixed-window schedulers leave 81.83% of cross-boundary cohort opportunity on the table.
· “Ready Cohorts: Bounding GPU Opportunity and Avoiding Host Round Trips in LLM-Agent Control”
Predict draft needs, don't explore: 41% faster edge LLMs
MemSpec keeps the best draft resident and prefetches the next, hitting 95-97% of an oracle schedule.
Velosiraptor turns hardware specs into verified OS memory code
One high-level specification yields addmap, setperms, delmap, and a simulator-side MMU, checked by an SMT solver.
Running app init as a small script cuts cold starts 1.67x
Shaves 100+ ms from cold starts by overlapping platform setup with app initialization.
· “Fast end-to-end cloud application cold-start with initscripts”
Forecast timestep becomes a dial
One weight set gives 6-hour detail for two weeks and 24-hour stability for month-long rollouts.
· “Timestep-Conditioned Transformers for Global Weather Forecasting”
One BPF map lookup blocks multi-hop microservice attacks
eMicro checks a request against its whole workflow history in the kernel, not just the last hop.
· “eMicro: Real-Time Multi-Hop Access Control for Microservices with eBPF”
Agentic AI strands CPU and GPU capacity on uniform servers
A production fleet study shows fragmented execution, and a prototype reclaims the slack while protecting tail latency.
AI framework generates datacenter policies that beat expert baselines
Autonomous policies beat expert baselines on placement, scaling, and power, cutting design effort from months to hours.
· “AtumAI: A Principled Framework for Agentic Generation of Datacenter Control-Plane Policies”
Seven commercial endpoint defenses fell to an autonomous AI system
Knowledge base and multi-agent pipeline hit 90% evasion on Windows Defender and 86.7% on Trend Micro.
· “Mutate to Bypass: Autonomous Endpoint Evasion via Knowledge-Driven Multi-Agent Orchestration”
Saving read positions can't guarantee exact Docker log recovery
In 60 runs, a position-persisting collector lost all retained logs whenever recovery required rediscovering an exited or restarted source.
One scanner misses a third of Docker Hub vulnerabilities
A six-scanner census of top images finds 96.3% carry a known flaw — and 66.8% of flaws are seen by only one tool.
· “Vulnerabilities, Secrets and Misconfiguration in the Highest-Exposure Docker Hub Images”
Own-the-machine filesystem checker catches a real Y2038 bug
Beneath the OS, an unmodified ext2 becomes its own oracle and the clock becomes a search axis.
· “Themis: A Filesystem Model Checker That Owns the Machine”
LLM pass rates collapse on file-system debugging and optimization
Across 505 tasks, the failure is incomplete edge-case reasoning and suboptimal code, not instruction following.
· “Benchmarking LLMs on File System Design and Implementation”
Three properties verify EDF schedulers in real RTOS code
Deductive checks on production kernels show deadline order holds under stated infrastructure assumptions
· “Deductive Verification for Earliest Deadline First Scheduler Implementations”
AI agents auto-write formal specs
Push-button TLA+ model checking over 48 projects replays each violation as a code-level test.
· “Specula: Scaling formal specifications for autonomous model checking of system code”
New language catches eBPF kernel/user mismatch at compile time
KernelScript unifies maps, program handles, and execution domains in one source, so silent runtime corruption becomes a compile error.
· “KernelScript: Cross-Boundary Typed DSL for eBPF Applications”
AI-native defined: autonomy is revision authority
A precise definition lets engineers and regulators tell real autonomy from a marketing label.
· “Defining AI-Native Systems: Autonomy as Revision Authority”
TRIM cuts AI patch bloat by up to 33 percent
Removing an agent's leftover search edits yields smaller, reviewable patches with no test-visible behavior change.
· “TRIM: Reducing AI-Generated CodeSlop via Agent Trajectory Minimization”
Fast-tracking low-priority lock holders cuts Android UI stalls 72%
SuperPass targets the runqueue wait that causes priority inversion, cutting janky frames by 29% with 0.74% CPU overhead.
· “SuperPass: Fast-Tracking Blocking Threads to Mitigate Priority Inversion on Mobile Devices”
Page-cache timing leaks cross container and VM boundaries
The deciding factor is storage: buffered shared files leak; direct I/O and dedicated block devices don't.
Workload-aware rollout boosts agentic RL throughput up to 1.6x
Synchronous RL stays on-policy while a hybrid of suffix reuse and cache-aware scheduling removes the rollout bottleneck.
· “WAR: Workload-Aware Rollouts for Synchronous Agentic Reinforcement Learning”
Kernel timing predicts colocation interference
By pairing models with offset kernel bursts, Roomie keeps latency violations near 10% under heavy load.
· “Roomie: Interference-Aware Colocation for Efficient Model Serving”
Scheduler converges to I/O limits within 5% under extreme skew
A three-stage governor decouples tenant budgets from hardware speed and topology in disaggregated storage.
· “Hardware-Transparent I/O Governance in Disaggregated Heterogeneous Storage”
Idle cores return power, not waste, on capped servers
Efficient waiting hands headroom back to active cores—so microsecond reallocation may cost more than it returns.
· “Rethinking Polling Efficiency in Service Core Network Stacks”
Multithreaded apps scale across CXL nodes without rewrites
Elastic pages and latency-driven placement deliver 1.5x-2.2x speedups over CXL-only memory and near-linear scaling.
· “Scaling Unmodified Multithreaded Applications with Elastic CXL-based Distributed Shared Memory”
Adding one select network to SIMD cuts low-bit GEMM latency 13.29x
Workload-aware support beats one fixed SIMD extension, especially for mixed-precision LLMs.
Fractional-bit CPU inference scales smoothly from 3 to 6 bits
A per-channel bit allocator plus compile-time layout makes fractional weights run on ordinary CPUs.
· “PolyQ: Codesigning End-to-End Quantization Framework for Scalable Edge CPU LLM Inference”
Kernel staging lifts large buffered reads 6.56x
MARS reorganizes page-cache work inside one read, reaching 36.87 GiB/s on five NVMe drives without changing read().
· “MARS: Multi-stage Accelerated Read Stack for Large-buffer Buffered Reads”
Inference becomes a remote OS call on dual-core MCUs in 15 µs
Lock-free shared rings and zero-copy tensors meet a 50 µs budget without doubling 64 KB RAM
· “Cross-Core Inference Offload as an Operating-System Service on Dual-Core Microcontrollers”
Inference pipelines become OS objects with fixed memory
Priority scheduling and per-frame arena reset keep microcontroller neural work heap-free and constant-footprint
Inference becomes a first-class Zephyr workload on MCUs with NPUs
Four subsystems cut fragmentation, state hygiene, and model lifecycle chores for edge AI
External index turns container cold starts from 20s to 2.8s
SOCI builds a seekable index over unmodified OCI images, letting containers fetch only the files they need at startup via HTTP range request
· “Seekable OCI: Lazy-Loading Container Images via Range-Request Indexing”
Pausing a distributed cluster for debugging without triggering timeouts
DDB embeds causality metadata in RPCs and virtualizes each process's clock, enabling GDB-style interactive debugging across 122-process分布式系统
· “DDB: Source-Level Interactive Debugging for Distributed Applications”
Complets make Arm POE2 safe to program
A Monitor-plus-trampoline model turns multi-index permission tables into three standard isolation patterns
LLM gang membership can change every token without breaking barriers
On real silicon, elastic cores beat static partitions at intermediate duty and return in 0.22 µs
One guest trip, N devices: 1.78× lower virtualized I/O latency
Moving multi-device fan-out into the SPDK backend replaces N stack traversals with one.
· “Cross-IP Request Coalescing: Relocating the Fan-out Point in Virtualized I/O”
eBPF fixes jump when you find the lost proof
Localizing where the verifier's safety proof disappeared lifts LLM repair by 11–21 points
· “Characterizing and Bridging the Diagnostic Gap in eBPF Verifier Rejections”
Portable C++ runtime deploys robot VLAs at 100% and 91%
Five layers unify multi-rate closed-loop inference across edge devices, robots, and simulators.
· “Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots”
Hide accelerator waits in tens of lines
Route the offload through the suspend/resume machinery every concurrent server already ships
· “Fine-Grained Computation Offload for Off-the-Shelf Servers in Tens of Lines”
GNSS pipeline timing becomes deterministic with logical time model
Federated execution logs confirm interrupt cadence, ring-buffer evolution, packetization, and physical-logical jitter.
· “Ensuring Deterministic Timing in a Federated GNSS Correction Pipeline with Lingua Franca”
Workflow combines CPU and GPU TEEs for private AI
End-to-end system protects models and data from cloud operators and Kubernetes admins while measuring overhead on TDX with H200.
Accessibility metadata powers AI agents without screenshot reliance
LUMOS turns native OS data into stable semantic blueprints with identifiers and actions for direct control.
· “LUMOS: A Semantic Operating-System Layer for Accessibility-Grounded AI Agents”
OS kernel engine enforces agent policies beyond tool calls
ActPlane uses an IFC DSL to catch indirect violations with 1.9-8.4 percent overhead.
· “ActPlane: Programmable OS-Level Policy Enforcement for Agent Harnesses”
Kops adds native ops to eBPF JIT via verified proof sequences
Each operation pairs a verifier-checked eBPF sequence with a native emit, delivering up to 24% microbenchmark gains while keeping TCB minima
· “Kops: Safely Extending the eBPF Compilation Pipeline with Native Operations”
Hierarchical CXL+RDMA pool speeds MicroVM restores 2.2x
Hot working-set pages sit in low-latency CXL while cold pages live in RDMA, with zero pages dropped entirely from the snapshot format.
· “Aquifer: Hierarchical Memory Pooling with CXL and RDMA for MicroVM Snapshots”
Master core LMS regulator cuts memory slowdowns over Memguard
Prediction from outside the regulated cores lowers contention effects on SPEC benchmarks by enforcing per-core bandwidth allocations.
· “LMS-AR: LMS Prediction-based Adaptive Regulator for Memory Bandwidth in Multicore Systems”
Android harness lifts AI agent success rate by 21 percent
AOHP adds three OS mechanisms so agents finish more tasks, spend fewer tokens, and respect security rules on existing Android devices.
· “AOHP: An Open-Source OS-Level Agent Harness for Personalized, Efficient and Secure Interaction”
Decoupled permissions speed secure mobile LLM inference 10x
Recallable isolation lets normal OS manage protected memory and NPU without reading contents.
On-device LLMs cut energy use up to 65% by lowering NPU speed
Model-structure predictions pick frequency settings that save power on phones, laptops, and boards while preserving response quality.
Reverse engineering maps Apple Neural Engine internals
Datapath, compiler format, weight compression, and command protocol detailed across A11 to A18 and M1 to M5 chips.
· “Apple Neural Engine: Architecture, Programming, and Performance”
OS reframes as intent filter for AI agents
Agents declare structured goals; the system builds constrained environments with mediation and auditing instead of granting direct resource
· “AgenticOS: An Intent-Oriented Secure Operating System Architecture for Autonomous AI Agents”
OS takes over LtRAM management with read-only app pages
Copy-on-write faults move wear-leveling and caching to the kernel, letting simpler DIMM hardware match DRAM speed on read-heavy tasks.
CloakLM fragments GPU weight memory to block exfiltration
Three software mechanisms break contiguous layouts so PCIe and HBM attacks see incoherent data, while inference runs at near-native speed.
· “CloakLM: Obfuscating GPU Memory Layout to Mitigate Model Ex-filtration for Serving”
Cordon wraps LLM tool calls in semantic transactions for pre-commit validation
A task-level boundary stages effects, tracks lineage, and validates the full flow so agents can roll back or audit before irreversible actio
Mixed-criticality on one connection: priority guaranteed
Sender-side scheme delivers critical messages 300 of 300 where FIFO delivers zero; QUIC with per-stream isolation is the escape.
io_uring requests become visible for under 10% cost on NVMe
One portable eBPF binary reconstructs every request and names the pathologies that own your p99.
· “uringscope: Portable, Low-Overhead Observability for io_uring”
Client-server system keeps language model music synced to live signals
High-frequency requests and clock-locked outputs enable real-time accompaniment across varying network delays, with quality tied to timing p
· “Real-Time Language Model Jamming: A Case Study for Live Music Accompaniment Generation”
Model virtualization shares one backbone to host 6x more tasks
FMplex cuts latency up to 80 percent versus separate instances by letting customized tasks share the same foundation model while keeping iso
· “FMplex: Model Virtualization for Serving Extensible Foundation Models”
Middleware gives microcontrollers per-container scheduling and access control
Metadata approach supports multiple runtimes on Cortex-M boards with at most 4 ms overhead per call.
Datalog language adds dynamic states to access policies
Authors encode fine-grained rules with process status and apply them to SELinux to test validity and expressiveness.
· “Policy Description Language for Authorization using Logic-Based Programming”
TOMOYO Linux bases access control on app execution history
Tracking state lets policies distinguish intended file accesses from malicious or erroneous ones.
· “TOMOYO Linux: A Mandatory Access Control Method Based on Application Execution State”
AgileOS virtualizes CUDA to protect GPU services
Client shims route calls to a trusted worker that owns the context and guards memory with PTX injection.
· “AgileOS: A GPU Operating System Layer for Protected CUDA Services”
Keeping old servers can cut total carbon versus upgrades
Lifecycle simulator finds manufacturing emissions outweigh efficiency gains under light loads or clean grids.
GPUs bypass CPU for 3.2x faster remote flash access
A GPU-centric stack lets GPUs start I/O to remote SSDs directly, cutting application times by 31 percent.
· “GNStor: Design of GPU-Native High-Performance Remote All-Flash Array”
Metastable failures trace to destabilizing component cycles
Identifying these cycles structurally lets systems avoid self-sustaining bad behavior even without overload.
Relative idleness ranking lifts agent throughput 20-71%
MORI places KV cache by ranking programs on a continuous busy-idle spectrum so GPU and CPU capacities stay matched on real coding traces.
Adaptive placement cuts robot task violations below 5%
Switching perception between robot and edge compute keeps tail latency low and CPU headroom healthy.
· “Edge-Based QoS-Aware Adaptive Task Placement: A Closed-Loop Control in Multi-Robot Systems”
Kernel fuzzer now sees argument values at function calls
LLVM pass records structured entry and return data via KCOV extension for value-aware mutation and root-cause analysis.
· “Beyond Edge Coverage: Per-Task Data-Flow Extraction at Kernel Function Boundaries via LLVM”
RTP-LLM speeds LLM loading by 4.7x-6.3x and cuts latency 35%
Disaggregated prefill-decode and hierarchical KV caching improve cache reuse and throughput over vLLM and SGLang in production traffic.
IORM manages I/O for thousands of consolidated databases without interference
It moves database semantics into storage scheduling so OLTP and analytical workloads share hardware with stable latency and no cross-tenant
· “IORM: Hierarchical I/O Governance for Thousands of Consolidated Databases on Oracle Exadata”
