arXiv subjects

Yusheng Zheng

Publications and source records attributed to Yusheng Zheng.

At least 37 records · Page 2Linked to original sources

arXiv2026

SysOM-AI: Continuous Cross-Layer Performance Diagnosis for Production AI Training

Performance diagnosis in production-scale AI training is challenging because subtle OS-level issues can trigger cascading GPU delays and network slowdowns, degrading training efficiency across thousands of GPUs. Existing profiling tools are limited to single system layers, incur prohibitive overhead (10--30%), or lack continuous deployment capabilities, resulting in manual analyses spanning days. We argue that continuous, cross-layer observability enabled by OS-level instrumentation and layered differential diagnosis is necessary to address this gap. We introduce SysOM-AI, a production observability system that continuously integrates CPU stack profiling, GPU kernel tracing, and NCCL event instrumentation via adaptive hybrid stack unwinding and eBPF-based tracing, incurring less than 0.4% overhead. Deployed at Alibaba across over 80,000 GPUs for more than one year, SysOM-AI helped diagnose 94 confirmed production issues, reducing median diagnosis time from days to approximately 10 minutes.

cs.PF↗

arXiv2026

ACRFence: Preventing Semantic Rollback Attacks in Agent Checkpoint-Restore

LLM agent frameworks increasingly offer checkpoint-restore for error recovery and exploration, advising developers to make external tool calls safe to retry. This advice assumes that a retried call will be identical to the original, an assumption that holds for traditional programs but fails for LLM agents, which re-synthesize subtly different requests after restore. Servers treat these re-generated requests as new, enabling duplicate payments, unauthorized reuse of consumed credentials, and other irreversible side effects; we term these semantic rollback attacks. We identify two attack classes, Action Replay and Authority Resurrection, validate them in a proof of concept experiment, and confirm that the problem has been independently acknowledged by framework maintainers. We propose ACRFence, a framework-agnostic mitigation that records irreversible tool effects and enforces replay-or-fork semantics upon restoration

cs.CR↗

arXiv2026

Fork, Explore, Commit: OS Primitives for Agentic Exploration

AI agents increasingly perform agentic exploration: pursuing multiple solution paths in parallel and committing only the successful one. Because each exploration path may modify files and spawn processes, agents require isolated environments with atomic commit and rollback semantics for both filesystem state and process state. We introduce the branch context, a new OS abstraction that provides: (1) copy-on-write state isolation with independent filesystem views and process groups, (2) a structured lifecycle of fork, explore, and commit/abort, (3) first-commit-wins resolution that automatically invalidates sibling branches, and (4) nestable contexts for hierarchical exploration. We realize branch contexts in Linux through two complementary components. First, BranchFS is a FUSE-based filesystem that gives each branch context an isolated copy-on-write workspace, with O(1) creation, atomic commit to the parent, and automatic sibling invalidation, all without root privileges. BranchFS is open sourced in https://github.com/multikernel/branchfs, along with a Python integration library, BranchContext, that provides ready-to-use agent exploration patterns. Second, branch() is a proposed Linux syscall that spawns processes into branch contexts with reliable termination, kernel-enforced sibling isolation, and first-commit-wins coordination. Preliminary evaluation of BranchFS shows sub-350 us branch creation independent of base filesystem size, and modification-proportional commit overhead (under 1 ms for small changes).

cs.OS↗

arXiv2026

SAIR: Cost-Efficient Multi-Stage ML Pipeline Autoscaling via In-Context Reinforcement Learning

Multi-stage ML inference pipelines are difficult to autoscale due to heterogeneous resources, cross-stage coupling, and dynamic bottleneck migration. We present SAIR, an autoscaling framework that uses an LLM as an in-context reinforcement learning controller, improving its policy online from reward-labeled interaction histories without gradient updates. SAIR combines Pareto-dominance reward shaping with a provable separation margin, surprisal-guided experience retrieval for context efficiency, and fine-grained GPU rate control via user-space CUDA interception. We provide regret analysis decomposing error into retrieval coverage and LLM selection components. On four ML serving pipelines under three workload patterns, SAIR achieves the best or tied-best P99 latency and effective resource cost among deployed baselines, improving P99 by up to 50% and reducing effective cost by up to 97% (under GPU rate-control assumptions), with 86% bottleneck detection accuracy and no offline training.

cs.LG↗

arXiv2026

Acceleration of Modelling with Physics Informed Learning: Frameworks and Perspectives for Real-Time Control of Electrochemical Devices

Electrochemical devices (batteries, fuel cells, and electrolyzers) are in full development, driven by the green energy transition. Their real-time control requires ms predictions in order to take critical decisions during fast transients or faults. The physics behind include coupled multi-physics phenomena that conventional finite element methods cannot solve so fast with the current CPU technology. This paper evaluates the potential of physics-informed machine learning represented by three frameworks: \ac{pinn}, \ac{pideeponet}, and \ac{pino} by evaluating their training effort, inference speed, and extrapolation capacity. Our analysis reveals valuable performance trade-offs. \acp{pinn} offer simplicity for fixed problem instances but require retraining for parameter changes. \ac{pideeponet} enables operator learning across varying conditions with mesh-free geometric flexibility. \ac{pino} delivers superior performance on regular grids, with the strongest extrapolation capabilities due to spectral derivative computation and resolution invariance. \ac{pideeponet} is particularly suited for irregular, unstructured geometries (e.g., porous electrodes or complex flow fields), while \ac{pino} works best for layered, structured-grid problems (e.g., transport across stacked electrochemical layers) requiring fast inference. Possible future applications include real-time lithium concentration prediction for safe fast-charging and micro short circuit detection, water management in fuel cells, and optimal power management in electrolyzers under intermittent renewable inputs. These findings establish physics-informed operator learning as a transformative approach for next-generation electrochemical device controller technology.

physics.app-ph↗

arXiv2026

Physics-informed neural network surrogate modeling of single particle model for lithium-ion batteries

Physics-based models play a key role in battery management, yet face challenges in real-time applications due to the high computational cost of solving coupled algebraic-partial differential equations. To accelerate model simulation, this study benchmarks three physics-informed neural network (PINN) architectures for modeling the battery single particle model, including two conventional PINN architectures and a DeepONet-based architecture. Both the accuracy and the generalization of these PINNs have been evaluated and compared under various current conditions. Our results highlight the potential of PINNs in modeling battery physics but also reveal limitations of conventional PINN architectures under highly dynamic current conditions. Among them, the Fourier-enhanced DeepONet achieves superior generalization performance and offers nearly a 10 times speedup compared with numerical solvers. This work provides an example of integrating physics-based models

physics.app-ph↗

arXiv2025

gpu_ext: Extensible OS Policies for GPUs via eBPF

Performance in modern GPU-centric systems increasingly depends on resource management policies, including memory placement, scheduling, and observability. However, uniform policies typically yield suboptimal performance across diverse workloads. Existing approaches present a tradeoff: user-space runtimes provide programmability and flexibility but lack cross-tenant visibility and fine-grained control of hardware resources; meanwhile, modifications to the OS kernel introduce significant complexity and safety risks. To address this, we argue that the GPU driver and device layer should provide an extensible OS interface for policy enforcement. While the emerging eBPF technology shows potential, directly applying existing host-side eBPF is insufficient because they lack visibility and control into critical device-side events, and directly embedding policy code into GPU kernels could compromise safety and efficiency. We propose gpu_ext, an eBPF-based runtime that treats the GPU driver and device as a programmable OS subsystem. gpu_ext extends GPU drivers by exposing safe programmable hooks and introduces a device-side eBPF runtime capable of executing verified policy logic within GPU kernels, enabling coherent and transparent policies. Evaluation across realistic workloads including inference, training, and vector search demonstrates that gpu_ext improves throughput by up to 4.8x and reduces tail latency by up to 2x, incurring low overhead, without modifying or restarting applications

cs.OS↗

arXiv2025

Towards Agentic OS: An LLM Agent Framework for Linux Schedulers

Operating system schedulers suffer from a fundamental semantic gap, where kernel policies fail to understand application-specific needs, leading to suboptimal performance. We introduce SchedCP, the first framework that enables fully autonomous Large Language Model (LLM) agents to safely and efficiently optimize Linux schedulers without human involvement. Our core insight is that the challenge is not merely to apply a better LLM, but to architect a decoupled control plane that separates the AI's role of semantic reasoning ("what to optimize") from the system's role of execution ("how to observe and act"), thereby separating the optimization problem into two stages: goal-inference and policy-synthesis. Implemented as Model Context Protocol(MCP) server, SchedCP provides a stable interface with three key services: a Workload Analysis Engine, an evolving Scheduler Policy Repository, and an Execution Verifier that validates all AI-generated code and configure before deployment with static and dynamic analysis. We demonstrate this architecture's power with sched-agent, a multi-agent system that autonomously analyzes workloads, synthesizes custom eBPF scheduling policies, and deploys them via the sched\_ext infrastructure. Our evaluation shows that SchedCP achieves up to an 1.79x performance improvement, and a 13x cost reduction compared to naive agentic approaches, all while maintaining high success rate. By bridging the semantic gap, SchedCP democratizes expert-level system optimization and represents a step towards creating truly self-optimizing, application-aware operating systems. The code is open-sourced in https://github.com/eunomia-bpf/schedcp

cs.AI↗

arXiv2025

MVVM: Deploy Your AI Agents-Securely, Efficiently, Everywhere

The rise of AI agents powered by Large Language Models (LLMs) presents critical challenges: how to securely execute and migrate these agents across heterogeneous environments while protecting sensitive user data, maintaining availability during network failures, minimizing response latency for time-critical decisions, and ensuring output safety in mission-critical applications. We present MVVM, a WebAssembly-based secure container framework that enables transparent live migration of LLM agent workspaces between edge devices and cloud servers with end-to-end privacy guarantees, resilient multi-tier replication, speculative execution for latency optimization, and integrated validation for safety assurance. MVVM introduces two key innovations: (1) a two-way sandboxing framework leveraging hardware enclaves and accelerator extensions that protects both the agent from malicious hosts and the host from compromised agents; (2) an efficient cross platform migration mechanism using WebAssembly and WASI's platform-agnostic design, enabling seamless movement across ARM phones, RISC-V MCUs, x86 servers, and heterogeneous accelerators; and three astonishing use cases: (1) privacy-aware daemon that automatically determines whether to execute locally or remotely based on data sensitivity and resource availability; (2) multi-tier replication with intelligent quality degradation that maintains service availability despite network failures or resource constraints; (3) a comprehensive execution framework combining speculative execution for 10x latency reduction with parallel validation that ensures output safety without compromising responsiveness. Our evaluation demonstrates that MVVM is validated on three separate devices across 18 workloads.

cs.OS↗

arXiv2025

Merging Physics-Based Synthetic Data and Machine Learning for Thermal Monitoring of Lithium-ion Batteries: The Role of Data Fidelity

Since the internal temperature is less accessible than surface temperature, there is an urgent need to develop accurate and real-time estimation algorithms for better thermal management and safety. This work presents a novel framework for resource-efficient and scalable development of accurate, robust, and adaptive internal temperature estimation algorithms by blending physics-based modeling with machine learning, in order to address the key challenges in data collection, model parameterization, and estimator design that traditionally hinder both approaches. In this framework, a physics-based model is leveraged to generate simulation data that includes different operating scenarios by sweeping the model parameters and input profiles. Such a cheap simulation dataset can be used to pre-train the machine learning algorithm to capture the underlying mapping relationship. To bridge the simulation-to-reality gap resulting from imperfect modeling, transfer learning with unsupervised domain adaptation is applied to fine-tune the pre-trained machine learning model, by using limited operational data (without internal temperature values) from target batteries. The proposed framework is validated under different operating conditions and across multiple cylindrical batteries with convective air cooling, achieving a root mean square error of 0.5 °C when relying solely on prior knowledge of battery thermal properties, and less than 0.1 °C when using thermal parameters close to the ground truth. Furthermore, the role of the simulation data quality in the proposed framework has been comprehensively investigated to identify promising ways of synthetic data generation to guarantee the performance of the machine learning model.

eess.SY↗

arXiv2025

CXLAimPod: CXL Memory is all you need in AI era

The proliferation of data-intensive applications, ranging from large language models to key-value stores, increasingly stresses memory systems with mixed read-write access patterns. Traditional half-duplex architectures such as DDR5 are ill-suited for such workloads, suffering bus turnaround penalties that reduce their effective bandwidth under mixed read-write patterns. Compute Express Link (CXL) offers a breakthrough with its full-duplex channels, yet this architectural potential remains untapped as existing software stacks are oblivious to this capability. This paper introduces CXLAimPod, an adaptive scheduling framework designed to bridge this software-hardware gap through system support, including cgroup-based hints for application-aware optimization. Our characterization quantifies the opportunity, revealing that CXL systems achieve 55-61% bandwidth improvement at balanced read-write ratios compared to flat DDR5 performance, demonstrating the benefits of full-duplex architecture. To realize this potential, the CXLAimPod framework integrates multiple scheduling strategies with a cgroup-based hint mechanism to navigate the trade-offs between throughput, latency, and overhead. Implemented efficiently within the Linux kernel via eBPF, CXLAimPod delivers significant performance improvements over default CXL configurations. Evaluation on diverse workloads shows 7.4% average improvement for Redis (with up to 150% for specific sequential patterns), 71.6% improvement for LLM text generation, and 9.1% for vector databases, demon-strating that duplex-aware scheduling can effectively exploit CXL's architectural advantages.

cs.OS↗

arXiv2025

AgentSight: System-Level Observability for AI Agents Using eBPF

Modern software infrastructure increasingly relies on LLM agents for development and maintenance, such as Claude Code and Gemini-cli. However, these AI agents differ fundamentally from traditional deterministic software, posing a significant challenge to conventional monitoring and debugging. This creates a critical semantic gap: existing tools observe either an agent's high-level intent (via LLM prompts) or its low-level actions (e.g., system calls), but cannot correlate these two views. This blindness makes it difficult to distinguish between benign operations, malicious attacks, and costly failures. We introduce AgentSight, an AgentOps observability framework that bridges this semantic gap using a hybrid approach. Our approach, boundary tracing, monitors agents from outside their application code at stable system interfaces using eBPF. AgentSight intercepts TLS-encrypted LLM traffic to extract semantic intent, monitors kernel events to observe system-wide effects, and causally correlates these two streams across process boundaries using a real-time engine and secondary LLM analysis. This instrumentation-free technique is framework-agnostic, resilient to rapid API changes, and incurs less than 3% performance overhead. Our evaluation shows AgentSight detects prompt injection attacks, identifies resource-wasting reasoning loops, and reveals hidden coordination bottlenecks in multi-agent systems. AgentSight is released as an open-source project at https://github.com/agent-sight/agentsight.

cs.OS↗

arXiv2025

HetGPU: The pursuit of making binary compatibility towards GPUs

Heterogeneous GPU infrastructures present a binary compatibility challenge: code compiled for one vendor's GPU will not run on another due to divergent instruction sets, execution models, and driver stacks . We propose hetGPU, a new system comprising a compiler, runtime, and abstraction layer that together enable a single GPU binary to execute on NVIDIA, AMD, Intel, and Tenstorrent hardware. The hetGPU compiler emits an architecture-agnostic GPU intermediate representation (IR) and inserts metadata for managing execution state. The hetGPU runtime then dynamically translates this IR to the target GPU's native code and provides a uniform abstraction of threads, memory, and synchronization. Our design tackles key challenges: differing SIMT vs. MIMD execution (warps on NVIDIA/AMD vs. many-core RISC-V on Tenstorrent), varied instruction sets, scheduling and memory model discrepancies, and the need for state serialization for live migration. We detail the hetGPU architecture, including the IR transformation pipeline, a state capture/reload mechanism for live GPU migration, and an abstraction layer that bridges warp-centric and core-centric designs. Preliminary evaluation demonstrates that unmodified GPU binaries compiled with hetGPU can be migrated across disparate GPUs with minimal overhead, opening the door to vendor-agnostic GPU computing.

cs.AR↗

arXiv2025

CXLMemSim: A pure software simulated CXL.mem for performance characterization

CXLMemSim is a fast, lightweight simulation framework that enables performance characterization of memory systems based on Compute Express Link (CXL) .mem technology. CXL.mem allows disaggregation and pooling of memory to mitigate memory stranding (underutilized memory trapped on fully loaded servers) in cloud and datacenter environments. However, CXL-attached memory introduces additional latency and bandwidth constraints compared to local DRAM, and real CXL .mem hardware is not yet widely available for empirical evaluation. CXLMemSim addresses this gap by attaching to unmodified applications and simulating CXL-based memory pools in software. It operates by tracing memory allocations and accesses using efficient kernel probes and hardware performance counters, dividing execution into epochs, and injecting timing delays to emulate various CXL .mem latency/bandwidth characteristics. This approach incurs modest runtime overhead while preserving realistic load/store memory access patterns. We implement CXLMemSim on commodity hardware without special devices, and our evaluation shows that it runs orders of magnitude faster than cycle-accurate simulators (e.g., Gem5) for real-world workloads, while accurately modeling the performance impact of CXL .mem. We demonstrate use cases where CXLMemSim enables experimentation with memory pooling configurations, scheduling policies, data migration strategies, and caching techniques that were previously infeasible to evaluate at scale. Key findings include the viability of software-based CXL .mem emulation with low overhead, insights into latency and congestion effects in memory pools, and guidance for system designers to optimize memory disaggregation. Overall, CXLMemSim provides a practical and extensible platform for researchers and practitioners to explore CXL.mem innovations before real hardware becomes commonplace.

cs.PF↗

arXiv2025

Exploiting Control-flow Enforcement Technology for Sound and Precise Static Binary Disassembly

Rewriting x86_64 binaries-whether for security hardening, dynamic instrumentation, or performance profiling is notoriously difficult due to variable-length instructions, interleaved code and data, and indirect jumps to arbitrary byte offsets. Existing solutions (e.g., "superset disassembly") ensure soundness but incur significant overhead and produce large rewritten binaries, especially for on-the-fly instrumentation. This paper addresses these challenges by introducing the Time Variance Authority (TVA), which leverages Intel's Control-Flow Enforcement Technology (CET). By recognizing endbr64 as the only valid indirect jump target, TVA prunes spurious disassembly paths while preserving soundness and emulates CET constraints on processors lacking native CET support, effectively mitigating ROP/JOP exploits without new hardware. We implement TVA by modernizing the Multiverse rewriter for 64-bit Linux. Our evaluation on SPEC CPU2017 and real-world applications shows that TVA-guided rewriting achieves up to 1.3x faster instrumentation time. These results underscore TVA's feasibility as a high-performance, uprobes-free alternative for robust x86_64 binary analysis and rewriting.

cs.AR↗

arXiv2024

Code-Survey: An LLM-Driven Methodology for Analyzing Large-Scale Codebases

Modern software systems like the Linux kernel are among the world's largest and most intricate codebases, continually evolving with new features and increasing complexity. Understanding these systems poses significant challenges due to their scale and the unstructured nature of development artifacts such as commits and mailing list discussions. We introduce Code-Survey, the first LLM-driven methodology designed to systematically explore and analyze large-scale codebases. The central principle behind Code-Survey is to treat LLMs as human participants, acknowledging that software development is also a social activity and thereby enabling the application of established social science techniques. By carefully designing surveys, Code-Survey transforms unstructured data, such as commits, emails, into organized, structured, and analyzable datasets. This enables quantitative analysis of complex software evolution and uncovers valuable insights related to design, implementation, maintenance, reliability, and security. To demonstrate the effectiveness of Code-Survey, we apply it to the Linux kernel's eBPF subsystem. We construct the Linux-bpf dataset, comprising over 670 features and 16,000 commits from the Linux community. Our quantitative analysis uncovers important insights into the evolution of eBPF, such as development patterns, feature interdependencies, and areas requiring attention for reliability and security. The insights have been initially validated by eBPF experts. Furthermore, Code-Survey can be directly applied to other subsystems within Linux and to other large-scale software projects. By providing a versatile tool for systematic analysis, Code-Survey facilitates a deeper understanding of complex software systems, enabling improvements across a variety of domains and supporting a wide range of empirical studies. The code and dataset is open-sourced.

cs.SE↗

arXiv2024

Wasm-bpf: Streamlining eBPF Deployment in Cloud Environments with WebAssembly

The extended Berkeley Packet Filter (eBPF) is extensively utilized for observability and performance analysis in cloud-native environments. However, deploying eBPF programs across a heterogeneous cloud environment presents challenges, including compatibility issues across different kernel versions, operating systems, runtimes, and architectures. Traditional deployment methods, such as standalone containers or tightly integrated core applications, are cumbersome and inefficient, particularly when dynamic plugin management is required. To address these challenges, we introduce Wasm-bpf, a lightweight runtime on WebAssembly and the WebAssembly System Interface (WASI). Leveraging Wasm platform independence and WASI standardized system interface, with enhanced relocation for different architectures, Wasm-bpf ensures cross-platform compatibility for eBPF programs. It simplifies deployment by integrating with container toolchains, allowing eBPF programs to be packaged as Wasm modules that can be easily managed within cloud environments. Additionally, Wasm-bpf supports dynamic plugin management in WebAssembly. Our implementation and evaluation demonstrate that Wasm-bpf introduces minimal overhead compared to native eBPF implementations while simplifying the deployment process.

cs.OS↗

arXiv2024

Novel Low-Complexity Model Development for Li-ion Cells Using Online Impedance Measurement

Modeling of Li-ion cells is used in battery management systems (BMS) to determine key states such as state-of-charge (SoC), state-of-health (SoH), etc. Accurate models are also useful in developing a cell-level digital-twin that can be used for protection and diagnostics in the BMS. In this paper, a low-complexity model development is proposed based on the equivalent circuit model (ECM) of the Li-ion cells. The proposed approach uses online impedance measurement at discrete frequencies to derive the ECM that matches closely with the results from the electro-impedance spectroscopy (EIS). The proposed method is suitable to be implemented in a microcontroller with low-computational power, typically used in BMS. Practical design guidelines are proposed to ensure fast and accurate model development. Using the proposed method to enhance the functions of a typical automotive BMS is described. Experimental validation is performed using large prismatic cells and small-capacity cylindrical cells. Root-mean-square error (RMSE) of less than 3\% is observed for a wide variation of operating conditions.

eess.SY↗