runtime/online scheduling

This work addresses the lack of a unified framework in current LLM agent workflows, which hinders method comparison and reproducibility. To resolve this, we propose the Agent Computation Graph (ACG) framework, which models workflows as computation graphs and adopts “structure determines timing” as a core principle. The framework explicitly distinguishes between reusable templates, runtime instance graphs, and execution traces, enabling a systematic categorization of static and dynamic optimization approaches. Through a comprehensive literature review and conceptual modeling, we develop a multidimensional evaluation framework that integrates structural properties, establishes precise terminology, and defines standardized evaluation criteria. This foundation supports a reproducible and highly comparable research paradigm for optimizing LLM agent workflows.

This paper addresses the lack of clarity regarding the diversity and evolutionary trajectories of modern workload schedulers. We propose a cross-layer taxonomy comprising three categories: OS process scheduling, cluster job scheduling, and big-data scheduling. Through algorithmic feature analysis and historical comparative study, we systematically characterize the design rationales, optimization objectives, and technological evolution of these schedulers, uncovering shared design patterns across local and distributed environments. Our key contribution is the first unified classification framework, which identifies three fundamental differentiating dimensions: resource abstraction granularity, scheduling timing, and feedback mechanism. Based on this analysis, we distill general-purpose scheduling design principles targeting heterogeneity, scalability, and QoS guarantees. The study provides both theoretical foundations and practical guidance for scheduler selection, cross-layer coordination optimization, and next-generation scheduler architecture design.

Traditional offline construction of Multi-Schedule Graphs (MSGs) fails to accommodate dynamic scenarios such as hardware failures and runtime slack variations, resulting in insufficient robustness of AI-driven scheduling in time-triggered systems. This paper proposes an online adaptive scheduling method based on reinforcement learning (RL), embedding an RL agent within a meta-scheduler to enable real-time MSG expansion, continuous scheduling policy optimization, and support for context-aware adaptation, mode switching, and dynamic performance-bound adjustment. Our key contribution lies in overcoming the limitations of offline training by enabling incremental, runtime construction of the MSG and concurrent exploration of scheduling policies. Experimental evaluation under strict deadline constraints demonstrates a 23.6% improvement in scheduling success rate and enhanced fault recovery capability; moreover, response latency to timing fluctuations and unexpected events is reduced by 41%, significantly improving system reliability and adaptability.

Modern OS schedulers rely on static, monolithic policies, struggling to simultaneously ensure fairness, throughput, and latency under heterogeneous hardware and dynamic, diverse workloads. This paper proposes the Adaptive Scheduling Agent (ASA) paradigm: it decouples offline modeling from online decision-making, employs lightweight ML models to identify abstract workload patterns, and dynamically selects the optimal specialized scheduler via time-weighted probabilistic voting and precomputed mapping tables. Implemented atop Linux’s sched_ext framework, ASA enables real-time, zero-training cross-platform policy switching. It introduces the first hybrid scheduling architecture that composes multiple expert schedulers on demand. Experiments show ASA outperforms EEVDF in 86.4% of user-perceived benchmark scenarios and selects a top-3 optimal scheduler in 78.6%—significantly enhancing overall scheduling efficacy.

Existing user-schedulable languages (USLs) struggle to reconcile fine-grained programmer control over scheduling logic with automated performance optimization. This paper proposes the Growable User-Schedulable Language (GUSL), enabling programmers to define novel scheduling operations externally to the compiler and compose trusted, fine-grained primitives into customizable scheduling libraries. We introduce Cursors—a novel mechanism that uniformly models three core extensibility elements: actions, predicates, and references—thereby achieving, for the first time, safe, user-driven language growth. Our approach integrates fine-grained primitive composition, Cursor abstraction, user-defined scheduling libraries, program transformation, and static analysis. Evaluation across 80+ high-performance kernels shows that GUSL reduces scheduling code size by an order of magnitude while matching the performance of hand-tuned implementations across x86, ARM, and RISC-V architectures.

HPC workloads are becoming increasingly heterogeneous, rendering traditional static heuristic schedulers inadequate for dynamic resource demands. To address this, we propose SchedTwin—the first real-time digital twin system for HPC job scheduling. It continuously ingests runtime event streams to drive high-fidelity discrete-event simulation, enabling rapid online evaluation of “what-if” scenarios across multiple scheduling policies and facilitating goal-driven, closed-loop adaptive scheduling. Deeply integrated with the PBS scheduler, SchedTwin achieves low-overhead (sub-10-second decision latency) and high-accuracy online policy optimization. Experimental evaluation in production environments demonstrates that SchedTwin significantly outperforms mainstream static schedulers—overcoming the longstanding dual bottlenecks of adaptability and timeliness inherent in conventional HPC scheduling approaches.

This work addresses the limitations of traditional large language model (LLM) serving systems, which rely on static, hand-crafted scheduling policies that struggle to adapt to runtime dynamics such as load fluctuations and cluster elasticity. To overcome this, the authors propose an online self-evolving system that leverages an LLM-driven program synthesis pipeline to continuously observe system states and autonomously rewrite serving policy code in real time. By transforming serving policies from fixed artifacts into “living code” continuously optimized by an LLM during deployment, this approach establishes a new paradigm for self-evolving LLM serving systems. Experimental results demonstrate that, across diverse dynamic scenarios, the proposed system achieves an average performance improvement of 34% over state-of-the-art baselines, with gains reaching up to 53%.

This work addresses the performance degradation of mixed multi-runtime and multi-process workloads under over-subscription, where traditional OS schedulers induce thread interference through periodic preemption, exacerbating lock contention and scalability collapse. To overcome this, the authors propose USF, a user-space scheduling framework that enables cross-process and multi-runtime cooperative scheduling without requiring privileged operations or application modifications. USF employs a cooperative policy, SCHED_COOP, which triggers context switches only when threads voluntarily block, thereby eliminating preemption-induced overheads. Built upon an extended GNU C library and the nOS-V runtime, USF maintains compatibility with mainstream parallel frameworks such as OpenMP. Evaluations on representative workloads—including nested BLAS, multi-process PyTorch with LLaMA-3 inference, and molecular dynamics simulations—demonstrate performance improvements of up to 2.4×.

This study addresses the challenge of objectively evaluating algorithm performance in the Dynamic Flexible Job Shop Scheduling Problem (DFJSP), which is hindered by reliance on static benchmarks and uncalibrated instance generators. To overcome this, the authors propose DynaSchedBench, a diagnostic framework featuring a Sequential Event Space Calibrator (SESC) that computes a Scheduling Stress Index (SSI) to enable controllable generation of problem instances with tunable difficulty. The framework supports snapshot-based simulation, agent testing, and visualization. It achieves, for the first time, precise and efficient control over DFJSP instance hardness. Empirical analysis reveals an “observability paradox” in large language model (LLM)-based scheduling agents—access to complete information unexpectedly degrades performance. Furthermore, most LLM agents exhibit only heuristic-level approximation capabilities, failing to surpass strong handcrafted heuristics, with limited gains from tool augmentation.

This study addresses the challenge of maintaining scheduling optimality in dynamic manufacturing environments, where frequent disruptions such as machine failures and incoming orders render existing methods ineffective for real-time adaptation. The authors propose a novel offline-online collaborative framework: in the offline phase, a diverse repository of high-quality scheduling rules is evolved within the MAP-Elites behavioral space using multi-role-guided initialization and topology-aware evolutionary operators; in the online phase, a probe-based state fingerprinting mechanism coupled with rapid forward simulation enables sub-second retrieval and deployment of the most suitable rule. Integrating large language model–driven heuristic design, the approach significantly outperforms state-of-the-art automated heuristics, classical dispatching rules, genetic programming, and deep reinforcement learning methods across 500 dynamic flexible job shop instances derived from real industrial data.

In dynamic multi-tenant environments, programmable caching engines such as CacheLib often suffer from performance degradation, memory inefficiency, and unfair service allocation due to rigid configuration schemes, insufficient runtime adaptability, and the absence of quality-of-service (QoS) guarantees. This work presents the first systematic empirical evaluation of CacheLib under fluctuating workloads across a range of configurations, uncovering its critical bottlenecks and limitations. The study not only quantifies the shortcomings of current designs in terms of fairness and efficiency but also provides clear guidance for future enhancements aimed at improving dynamic adaptability, QoS support, and programmability in caching systems.