In serial batch (s-batch) scheduling, jobs are grouped in batches and processed sequentially within their batch. This paper considers multiple parallel machines, nonidentical job weights and release times, and sequence-dependent setup times between batches of different families. Although s-batch has been widely studied in the literature, very few papers have taken into account a minimum batch size, typical in practical settings such as semiconductor manufacturing and the metal industry. The problem with this minimum batch size requirement has been mostly tackled with dynamic programming and meta-heuristics, and no article has ever used constraint programming (CP) to do so. This paper fills this gap by proposing, for the first time, a CP model for s-batching with minimum batch size. The computational experiments on standard cases compare the CP model with two existing mixed-integer programming (MIP) models from the literature. The results demonstrate the versatility of the proposed CP model to handle multiple variations of s-batching; and its ability to produce, in large instances, better solutions than the MIP models faster.
Dynamic scheduling in real-world environments often struggles to adapt to unforeseen disruptions, making traditional static scheduling methods and human-designed heuristics inadequate. This paper introduces an innovative approach that combines Genetic Programming (GP) with a Transformer trained through Reinforcement Learning (GPRT), specifically designed to tackle the complexities of dynamic scheduling scenarios. GPRT leverages the Transformer to refine heuristics generated by GP while also seeding and guiding the evolution of GP. This dual functionality enhances the adaptability and effectiveness of the scheduling heuristics, enabling them to better respond to the dynamic nature of real-world tasks. The efficacy of this integrated approach is demonstrated through a practical application in container terminal truck scheduling, where the GPRT method outperforms traditional GP, standalone Transformer methods, and other state-of-the-art competitors. The key contribution of this research is the development of the GPRT method, which showcases a novel combination of GP and Reinforcement Learning (RL) to produce robust and efficient scheduling solutions. Importantly, GPRT is not limited to container port truck scheduling; it offers a versatile framework applicable to various dynamic scheduling challenges. Its practicality, coupled with its interpretability and ease of modification, makes it a valuable tool for diverse real-world scenarios.
In realistic production scenarios, Advanced Planning and Scheduling (APS) tools often require manual intervention by production planners, as the system works with incomplete information, resulting in suboptimal schedules. Often, the preferable solution is not found just because of the too-restrictive constraints specifying the optimization problem, representing bottlenecks in the schedule. To provide computer-assisted support for decision-making, we aim to automatically identify bottlenecks in the given schedule while linking them to the particular constraints to be relaxed. In this work, we address the problem of reducing the tardiness of a particular project in an obtained schedule in the resource-constrained project scheduling problem by relaxing constraints related to identified bottlenecks. We develop two methods for this purpose. The first method adapts existing approaches from the job shop literature and utilizes them for so-called untargeted relaxations. The second method identifies potential improvements in relaxed versions of the problem and proposes targeted relaxations. Surprisingly, the untargeted relaxations result in improvements comparable to the targeted relaxations.
Large language model (LLM) inference serving systems are essential to various LLM-based applications. As demand for LLM services continues to grow, scaling these systems to handle high request rates while meeting latency Service-Level Objectives (SLOs), referred to as effective throughput, becomes critical. However, existing systems often struggle to improve effective throughput, primarily due to a significant decline in Time To First Token (TTFT) SLO attainment. We identify two major causes of this bottleneck: (1) memory-intensive KV cache that limits batch size expansion under GPU memory constraints, and (2) rigid batch composition enforced by the default First-Come-First-Serve scheduling policy. In this paper, we introduce Apt-Serve, a scalable framework designed to enhance effective throughput in LLM inference serving. Apt-Serve features a new hybrid cache scheme that combines KV cache with a memory-efficient hidden cache for reusable input hidden state vectors, allowing large batch sizes and improving request concurrency. Based on the hybrid cache, Apt-Serve employs an adaptive runtime scheduling mechanism that dynamically optimizes batch composition. We formally define the adaptive scheduling optimization problem and propose an efficient algorithm with theoretical guarantees. Extensive evaluations on three real-world datasets and LLMs ranging from 13B to 66B parameters demonstrate that Apt-Serve achieves up to 8.8x improvement in effective throughput compared to the state-of-the-art inference serving systems.
As demand for Large Language Models (LLMs) and AI agents rapidly grows, optimizing systems for efficient LLM inference becomes critical. While significant efforts have focused on system-level engineering, little is explored from a mathematical modeling and queuing perspective. In this paper, we aim to develop the queuing fundamentals for large language model (LLM) inference, bridging the gap between the queueing theory and LLM system communities. In particular, we study the throughput aspect in LLM inference systems. We prove that a large class of 'work-conserving' scheduling algorithms can achieve maximum throughput for individual inference LLM engine, highlighting 'work-conserving' as a key design principle in practice. In a network of LLM agents, work-conserving scheduling alone is insufficient, particularly when facing specific workload structures and multi-class workflows that require more sophisticated scheduling strategies. Evaluations of real-world systems show that Orca and Sarathi-serve are throughput-optimal, reassuring practitioners, while FasterTransformer and vanilla vLLM are not maximally stable and should be used with caution. Our results highlight the substantial benefits that the queueing community can offer in improving LLM inference systems and call for more interdisciplinary development.
In this paper, we investigate a joint source-channel encoding (JSCE) scheme in an intelligent reflecting surface (IRS)-assisted multi-user semantic communication system. Semantic encoding not only compresses redundant information, but also enhances information orthogonality in a semantic feature space. Meanwhile, the IRS can adjust the spatial orthogonality, enabling concurrent multi-user semantic communication in densely deployed wireless networks to improve spectrum efficiency. We aim to maximize the users' semantic throughput by jointly optimizing the users' scheduling, the IRS's passive beamforming, and the semantic encoding strategies. To tackle this non-convex problem, we propose an explainable deep neural network-driven deep reinforcement learning (XD-DRL) framework. Specifically, we employ a deep neural network (DNN) to serve as a joint source-channel semantic encoder, enabling transmitters to extract semantic features from raw images. By leveraging structural similarity, we assign some DNN weight coefficients as the IRS's phase shifts, allowing simultaneous optimization of IRS's passive beamforming and DNN training. Given the IRS's passive beamforming and semantic encoding strategies, user scheduling is optimized using the DRL method. Numerical results validate that our JSCE scheme achieves superior semantic throughput compared to the conventional schemes and efficiently reduces the semantic encoder's mode size in multi-user scenarios.
Open RAN (O-RAN) fosters multi-vendor interoperability and data-driven control but simultaneously introduces the challenge of coordinating pre-trained xApps that may produce conflicting actions. Although O-RAN specifications mandate offline training and validation to prevent the deployment of untrained or inadequately tested models, operational conflicts can still arise under dynamic and context-dependent conditions.This work proposes a scheduler-based conflict mitigation framework to address these challenges without requiring training xApps together or further xApp re-training. By examining an indirect conflict involving power and resource block allocation xApps and employing an Advantage Actor-Critic (A2C) approach to train both xApps and the scheduler, we illustrate that a straightforward A2C-based scheduler improves performance relative to independently deployed xApps and conflicting cases. Notably, among all tested deployment scenarios (including individual xApp deployment, multiple conflicting xApps, and limited scheduler configurations), augmenting the system with baseline xApps and enabling the scheduler to select from a broader pool achieves the highest total transmission rate, thereby underscoring the importance of adaptive scheduling mechanisms. These findings highlight the context-dependent nature of conflicts in automated network management, as two xApps may conflict under certain conditions but coexist under others. Consequently, the ability to dynamically update and adapt the scheduler to accommodate diverse operational intents is vital for future network deployments. By offering dynamic scheduling without re-training xApps, this framework advances practical conflict resolution solutions while supporting real-world scalability.
This research presents a novel framework integrating Flexible-Duplex (FlexD) and Integrated Sensing and Communications (ISAC) technologies to address the challenges of spectrum efficiency and resource optimization in next-generation wireless networks. We develop a unified system model for a dual-functional radar-communication base station with multiple-input multiple-output capabilities, enabling dynamic uplink and downlink channel allocation. The framework maximizes network throughput while maintaining radar sensing performance, subject to signal-to-clutter-plus-noise ratio (SCNR) requirements and power constraints. Given the non-convex and combinatorial nature of the resulting optimization problem, we propose an iterative algorithm that converges to a locally optimal solution. Extensive simulations demonstrate the superiority of the proposed FlexD-ISAC framework compared to conventional half-duplex networks. Additionally, sensitivity analyses reveal the impact of SCNR requirements and power constraints on system performance, providing valuable insights for practical implementation. This work establishes a foundation for future research in dynamic, resource-efficient wireless systems that simultaneously support sensing and communication capabilities.
Fault tolerance in Deep Neural Networks (DNNs) deployed on resource-constrained systems presents unique challenges for high-accuracy applications with strict timing requirements. Memory bit-flips can severely degrade DNN accuracy, while traditional protection approaches like Triple Modular Redundancy (TMR) often sacrifice accuracy to maintain reliability, creating a three-way dilemma between reliability, accuracy, and timeliness. We introduce NAPER, a novel protection approach that addresses this challenge through ensemble learning. Unlike conventional redundancy methods, NAPER employs heterogeneous model redundancy, where diverse models collectively achieve higher accuracy than any individual model. This is complemented by an efficient fault detection mechanism and a real-time scheduler that prioritizes meeting deadlines by intelligently scheduling recovery operations without interrupting inference. Our evaluations demonstrate NAPER's superiority: 40% faster inference in both normal and fault conditions, maintained accuracy 4.2% higher than TMR-based strategies, and guaranteed uninterrupted operation even during fault recovery. NAPER effectively balances the competing demands of accuracy, reliability, and timeliness in real-time DNN applications
This work focuses on minimizing the age of information for multiple energy harvesting sources that sample data and transmit it to a sink node. At each time, the central scheduler selects one of the sources to probe the quality of its channel to the sink node, and then the assessed channel quality is utilized to determine whether a source will sample and send the packet. For a single source case, we assume that the probed channel quality is known at each time instant, model the problem of AoI minimization as a Markov decision process, and prove the optimal sampling policy threshold structure. We then use this threshold structure and propose an AEC-SW-UCRL2 algorithm to handle unknown and time varying energy harvesting rate and channel statistics, motivated by the popular SWUCRL2 algorithm for non stationary reinforcement learning. This algorithm is applicable when an upper bound is available for the total variation of each of these quantities over a time horizon. Furthermore, in situations where these variation budgets are not accessible, we introduce the AEC-BORL algorithm, motivated by the well known BORL algorithm. For the multiple source case, we demonstrate that the AoI minimization problem can be formulated as a constrained MDP, which can be relaxed using a Lagrange multiplier and decoupled into sub problems across source nodes. We also derive Whittle index based source scheduling policy for probing and an optimal threshold policy for source sampling. We next leverage this Whittle index and threshold structure to develop the WIT-SW-UCRL2 algorithm for unknown time varying energy harvesting rates and channel statistics under their respective variation budgets. Moreover, we also proposed a Whittle index and threshold based bandit over reinforcement learning (WIT-BORL) algorithm for unknown variation budgets. Finally, we numerically demonstrate the efficacy of our algorithms.
Efficient scheduling remains a critical challenge in various domains, requiring solutions to complex NP-hard optimization problems to achieve optimal resource allocation and maximize productivity. In this paper, we introduce a framework called Transformer-Based Task Scheduling System (TRATSS), designed to address the intricacies of single agent scheduling in graph-based environments. By integrating the latest advancements in reinforcement learning and transformer architecture, TRATSS provides a novel system that outputs optimized task scheduling decisions while dynamically adapting to evolving task requirements and resource availability. Leveraging the self-attention mechanism in transformers, TRATSS effectively captures complex task dependencies, thereby providing solutions with enhanced resource utilization and task completion efficiency. Experimental evaluations on benchmark datasets demonstrate TRATSS's effectiveness in providing high-quality solutions to scheduling problems that involve multiple action profiles.
In the forthcoming 6G era, extend reality (XR) has been regarded as an emerging application for ultra-reliable and low latency communications (URLLC) with new traffic characteristics and more stringent requirements. In addition to the quasi-periodical traffic in XR, burst traffic with both large frame size and random arrivals in some real world low latency communication scenarios has become the leading cause of network congestion or even collapse, and there still lacks an efficient algorithm for the resource scheduling problem under burst traffic with hard latency constraints. We propose a novel hybrid reinforcement learning framework for resource scheduling with hard latency constraints (HRL-RSHLC), which reuses polices from both old policies learned under other similar environments and domain-knowledge-based (DK) policies constructed using expert knowledge to improve the performance. The joint optimization of the policy reuse probabilities and new policy is formulated as an Markov Decision Problem (MDP), which maximizes the hard-latency constrained effective throughput (HLC-ET) of users. We prove that the proposed HRL-RSHLC can converge to KKT points with an arbitrary initial point. Simulations show that HRL-RSHLC can achieve superior performance with faster convergence speed compared to baseline algorithms.
AI agents are autonomous systems that can execute specific tasks based on predefined programming. Here, we present SkillFlow, a modular, technology-agnostic framework that allows agents to expand their functionality in an ad-hoc fashion by acquiring new skills from their environment or other agents. We present a theoretical model that examines under which conditions this framework would be beneficial, and we then explore SkillFlow's ability to accelerate task completion and lead to lower cumulative costs in a real-world application, namely scheduling agents for calendar events. We demonstrate that within a few iterations, SkillFlow leads to considerable (24.8%, p-value = $6.4\times10^{-3}$) gains in time and cost, especially when the communication cost is high. Finally, we draw analogies from well-studied biological systems and compare this framework to that of lateral gene transfer, a significant process of adaptation and evolution in novel environments.
Traditional 3D content representations include dense point clouds that consume large amounts of data and hence network bandwidth, while newer representations such as neural radiance fields suffer from poor frame rates due to their non-standard volumetric rendering pipeline. 3D Gaussian splats (3DGS) can be seen as a generalization of point clouds that meet the best of both worlds, with high visual quality and efficient rendering for real-time frame rates. However, delivering 3DGS scenes from a hosting server to client devices is still challenging due to high network data consumption (e.g., 1.5 GB for a single scene). The goal of this work is to create an efficient 3D content delivery framework that allows users to view high quality 3D scenes with 3DGS as the underlying data representation. The main contributions of the paper are: (1) Creating new layered 3DGS scenes for efficient delivery, (2) Scheduling algorithms to choose what splats to download at what time, and (3) Trace-driven experiments from users wearing virtual reality headsets to evaluate the visual quality and latency. Our system for Layered 3D Gaussian Splats delivery L3GS demonstrates high visual quality, achieving 16.9% higher average SSIM compared to baselines, and also works with other compressed 3DGS representations.
This paper introduces a novel reinforcement learning (RL) approach to scheduling mixed-criticality (MC) systems on processors with varying speeds. Building upon the foundation laid by [1], we extend their work to address the non-preemptive scheduling problem, which is known to be NP-hard. By modeling this scheduling challenge as a Markov Decision Process (MDP), we develop an RL agent capable of generating near-optimal schedules for real-time MC systems. Our RL-based scheduler prioritizes high-critical tasks while maintaining overall system performance. Through extensive experiments, we demonstrate the scalability and effectiveness of our approach. The RL scheduler significantly improves task completion rates, achieving around 80% overall and 85% for high-criticality tasks across 100,000 instances of synthetic data and real data under varying system conditions. Moreover, under stable conditions without degradation, the scheduler achieves 94% overall task completion and 93% for high-criticality tasks. These results highlight the potential of RL-based schedulers in real-time and safety-critical applications, offering substantial improvements in handling complex and dynamic scheduling scenarios.
The Mixture of Experts (MoE) architecture has demonstrated significant advantages as it enables to increase the model capacity without a proportional increase in computation. However, the large MoE model size still introduces substantial memory demands, which usually requires expert offloading on resource-constrained platforms and incurs significant overhead. Hybrid CPU-GPU inference has been proposed to leverage CPU computation to reduce expert loading overhead but faces major challenges: on one hand, the expert activation patterns of MoE models are highly unstable, rendering the fixed mapping strategies in existing works inefficient; on the other hand, the hybrid CPU-GPU schedule for MoE is inherently complex due to the diverse expert sizes, structures, uneven workload distribution, etc. To address these challenges, in this paper, we propose HybriMoE, a hybrid CPU-GPU inference framework that improves resource utilization through a novel CPU-GPU scheduling and cache management system. HybriMoE introduces (i) a dynamic intra-layer scheduling strategy to balance workloads across CPU and GPU, (ii) an impact-driven inter-layer prefetching algorithm, and (iii) a score-based caching algorithm to mitigate expert activation instability. We implement HybriMoE on top of the kTransformers framework and evaluate it on three widely used MoE-based LLMs. Experimental results demonstrate that HybriMoE achieves an average speedup of 1.33$\times$ in the prefill stage and 1.70$\times$ in the decode stage compared to state-of-the-art hybrid MoE inference framework. Our code is available at: https://github.com/PKU-SEC-Lab/HybriMoE.
The benefits of Deep Learning (DL) impose significant pressure on GPU resources, particularly within GPU cluster, where Out-Of-Memory (OOM) errors present a primary impediment to model training and efficient resource utilization. Conventional OOM estimation techniques, relying either on static graph analysis or direct GPU memory profiling, suffer from inherent limitations: static analysis often fails to capture model dynamics, whereas GPU-based profiling intensifies contention for scarce GPU resources. To overcome these constraints, VeritasEst emerges. It is an innovative, entirely CPU-based analysis tool capable of accurately predicting the peak GPU memory required for DL training tasks without accessing the target GPU. This "offline" prediction capability is core advantage of VeritasEst, allowing accurate memory footprint information to be obtained before task scheduling, thereby effectively preventing OOM and optimizing GPU allocation. Its performance was validated through thousands of experimental runs across convolutional neural network (CNN) models: Compared to baseline GPU memory estimators, VeritasEst significantly reduces the relative error by 84% and lowers the estimation failure probability by 73%. VeritasEst represents a key step towards efficient and predictable DL training in resource-constrained environments.
Disaggregated inference has become an essential framework that separates the prefill (P) and decode (D) stages in large language model inference to improve throughput. However, the KV cache transfer faces significant delays between prefill and decode nodes. The block-wise calling method and discontinuous KV cache memory allocation increase the number of calls to the transmission kernel. Additionally, existing frameworks often fix the roles of P and D nodes, leading to computational imbalances. In this paper, we propose FlowKV, a novel disaggregated inference framework, which reduces the average transmission latency of KV cache by 96%, from 0.944s to 0.053s, almost eliminating the transfer time relative to the total request latency by optimizing the KV cache transfer. FlowKV introduces the Load-Aware Scheduler for balanced request scheduling and flexible PD node allocation. This design maximizes hardware resource utilization, achieving peak system throughput across various scenarios, including normal, computational imbalance, and extreme overload conditions. Experimental results demonstrate that FlowKV significantly accelerates inference by 15.2%-48.9% on LongBench dataset compared to the baseline and supports applications with heterogeneous GPUs.
With the rapid expansion of cloud computing applications, optimizing resource allocation has become crucial for improving system performance and cost efficiency. This paper proposes an intelligent resource allocation algorithm that leverages deep learning (LSTM) for demand prediction and reinforcement learning (DQN) for dynamic scheduling. By accurately forecasting computing resource demands and enabling real-time adjustments, the proposed system enhances resource utilization by 32.5%, reduces average response time by 43.3%, and lowers operational costs by 26.6%. Experimental results in a production cloud environment confirm that the method significantly improves efficiency while maintaining high service quality. This study provides a scalable and effective solution for intelligent cloud resource management, offering valuable insights for future cloud optimization strategies.
The high memory and computation demand of large language models (LLMs) makes them challenging to be deployed on consumer devices due to limited GPU memory. Offloading can mitigate the memory constraint but often suffers from low GPU utilization, leading to low inference efficiency. In this work, we propose a novel framework, called pipelined offloading (PIPO), for efficient inference on consumer devices. PIPO designs a fine-grained offloading pipeline, complemented with optimized data transfer and computation, to achieve high concurrency and efficient scheduling for inference. Experimental results show that compared with state-of-the-art baseline, PIPO increases GPU utilization from below 40% to over 90% and achieves up to 3.1$\times$ higher throughput, running on a laptop equipped with a RTX3060 GPU of 6GB memory.