Cloud-edge environments require efficient task scheduling and resource management as they serve as key factors supporting the diverse, dynamic, and latency sensitive workloads of the modern smart infrastructures. Nevertheless, the non-uniformity of work, inconsistent work loads, and cold-start delays are major challenges to scalable and sustainable scheduling of tasks. In this paper, the authors introduce a new Context-Aware Predictive Reinforcement Framework (CAPRF++) of generalized task scheduling under unobservable workload scenarios, which is developed to be able to manage multi-modal workloads in distributed cloud-edge systems. In order to be compatible with heterogeneous forms of workload, a Context-Aware Task Feature Construction pipeline is suggested to encode metadata including task type, compute demand, deadline, and sustainability constraints into structured forms. Resource indexing strategy based on B+ tree can be used to achieve rapid and scalable task-device matching using real-time performance and energy indicators. It presents a new Weighted Multi-Objective PPO scheduling policy using a Weighted Tradeoff Controller that limits the policy choice to feasible SLA-energy-latency tradeoff space. Moreover, a FIFO replay buffer-based Adaptive Trajectory-Buffered PPO algorithm, in combination with a multi-objective reward model can be used to perform robust policy learning in changing workloads, balancing SLA compliance, execution time, energy consumption, and carbon footprint. In order to reduce cold-start overheads, a Predictive Prewarming module will be implemented which will pre- initialize runtime containers according to the workload expectations. The entire assessment is done with a massive simulated cloudedge condition that attempts to simulate actual workload dynamics and system heterogeneity. Significant simulation-based analyses show that CAPRF++ is an efficient system in terms of task completion latency, SLA satisfaction, energy consumption and generalization to unseen workloads in dynamic environments.
The design and verification of safety-critical real-time systems rely on schedulability analysis. Exact tests may provide correctness guarantees, but they are often scheduler specific, computationally intractable at high scale, and pessimistic when used in iterative analysis pipelines. The paper introduces the SAFE-TSFormer algorithm to handle periodic real-time task-sets of variable cardinality under multiple uniprocessor scheduling policies. The SAFE-TSFormer approach employs a two-stage strategy. At the outset, it performs non-linear feature encoding at the task level, generating encoded representations for each task. These representations are then processed using transformer-based sequence modeling with masked attention to capture inter-task interactions. SAFE-TSFormer is designed as a conservative screening mechanism, prioritizing the suppression of false positives through a safety-aware training objective. We present results from experiments conducted on a large set of inflated periodic real-time task sets comprising over 3, 60, 000 instances. The proposed model generalizes to task-set sizes (number of real-time tasks) significantly higher than those observed during training. Specifically, on entirely new sets of tasks containing up to thirty-two tasks, SAFE-TSFormer’s overall accuracy reaches 96.56% . The results demonstrate the ability of a learner-based schedulability predictor to operate as a complementary technology within a solution optimization system, which rapidly and scalably screens candidates for feasibility without sacrificing safety, with high accuracy ( 98.98% ) and very low false positive rates ( <0.4% ).
We present the first (to our knowledge) Deep-Learning based framework for real-time schedulability-analysis that guarantees to never incorrectly mis-classify an unschedulable system as being schedulable, and is hence suitable for use in safety-critical scenarios. We relate applicability of this framework to well-understood concepts in computational complexity theory: membership in the complexity class NP. We apply the framework upon the widely-studied schedulability analysis problems of determining whether a given constrained-deadline sporadic task system is schedulable on a preemptive uniprocessor under both Deadline-Monotonic and EDF scheduling. As a proof-of-concept, we implement our framework for Deadline-Monotonic scheduling, and demonstrate that it has a predictive accuracy exceeding 70% for systems of as many as 20 tasks without making any unsafe predictions. Furthermore, the implementation has very small ( <1 ms on two widely-used embedded platforms; <4 s on an embedded FPGA) and highly predictable running times.
In scheduling real-time tasks, we face the challenge of meeting hard deadlines while optimizing for some other objective, such as minimizing energy consumption. Formulating the optimization as a Multi-Armed Bandit (MAB) problem allows us to use MAB strategies to balance the exploitation of good choices based on observed data with the exploration of potentially better options. In this paper, we integrate hard real-time constraints with MAB strategies for resource management of a Stochastic Parallel Synchronous Task. On a platform with M cores available for the task, m≤ M cores are initially assigned. Prior work has shown how to compute a virtual deadline such that assigning all M cores to the task if it has not completed by this virtual deadline guarantees that the deadline will be met. An MAB strategy is used to select the value of m . A Dynamic Power Management (DPM) energy model considering CPU sockets and sleep states is described. Experimental evaluation shows that MAB strategies learn consistently suitable m , and perform well compared to binary exponential search and greedy methods.
Machine learning (ML) is increasingly being integrated into real-time embedded systems, enabling intelligent decision-making in applications such as autonomous driving and industrial automation. However, ensuring predictable execution of deep neural network (DNN) inference remains a major challenge, as real-time systems must meet strict timing constraints to guarantee safety and reliability. This paper identifies key challenges in achieving real-time AI inference in embedded systems, including limited memory capacity, high energy consumption, efficient multi-DNN scheduling, and heterogeneous resource management. To address these challenges, we emphasize the need for advanced scheduling algorithms to efficiently allocate heterogeneous computing resources across multiple DNNs, hierarchical memory management to reduce memory bottlenecks, and real-time neural architecture search and optimization techniques to enhance AI model performance under strict timing constraints. Furthermore, we discuss future research directions aimed at improving real-time AI execution, including time-predictable scheduling frameworks to ensure consistent inference latency, cross-device AI workload management to optimize resource utilization across heterogeneous processors, and benchmarking methodologies to systematically evaluate performance, timing guarantees, and energy efficiency in real-time AI systems. Advancing these research areas will enhance the reliability, efficiency, and scalability of AI-driven embedded systems, bridging the gap between ML advancements and real-time system requirements.
Many real-time problems can be characterized as combinatorial optimization problems where exact solutions are infeasible at scale. As problem complexity grows, handcrafted heuristics become increasingly difficult to design. Reinforcement learning (RL) has emerged as a promising alternative, enabling the discovery of decision-making policies without requiring explicit supervision. While RL does not guarantee optimality, it provides adaptive heuristics to solve complex problems. This paper explores the potential of RL for real-time resource management, outlining key principles, demonstrating an application to directed acyclic graph (DAG) scheduling, and identifying open challenges for future research.
The embedding of AI algorithms in critical systems faces significant scientific issues, among these issues, guaranteeing the temporal determinism of AI algorithms. Real-time scheduling theory can help addressing this challenge but necessitate the proposition of new real-time task models and scheduling algorithms. In this context, we identify three problems to address, firstly, the proposition of new task models for AI algorithms that depends on the class of AI algorithms to which they belong, secondly, how to integrate the architecture characteristics into the real-time task model, and finally, the establishment of benchmarks that are specific to the evaluation of the execution of AI algorithms with real-time constraints. To illustrate our statement, in this work we focus on the problem of the implementation of convolutional neural networks (CNNs) during the inference phase executed on NVIDIA GPUs.
We propose two complementary research directions, “Time for ML” and “ML for Time”, that we believe to be critical for the deployment of machine-learning (ML) applications in time-sensitive applications. “Time for ML” refers to ML systems that are aware of and can adapt to dynamic time constraints regarding their execution, while “ML for Time” refers to ML systems that are aware of and can deal with data’s temporal aspects, such as misalignment. We believe these two directions are complementary and can be combined to provide more robust and reliable machine learning systems.
Reinforcement Learning (RL) has emerged as a promising tool for decision-making in various applications, particularly in uncertain environments. While its adoption in embedded systems—especially hard real-time systems—faces challenges due to stringent timing constraints, integrating shielding mechanisms may offer a pathway for RL to optimize its scheduling decisions, preserving worst-case timing guarantees. This position paper shows a use case where RL selects compliant execution versions for fault-tolerant real-time systems while minimizing the system utilization in runtime. Furthermore, we discuss possible directions for further exploring RL’s role in real-time systems for improved adaptability.
Recent advances in AI culminate a shift in science and engineering away from strong reliance on algorithmic and symbolic knowledge towards new data-driven approaches. How does the emerging intelligent data-centric world impact research on real-time and embedded computing? We argue for two effects: (1) new challenges in embedded system contexts, and (2) new opportunities for community expansion beyond the embedded domain. First, on the embedded system side, the shifting nature of computing towards data-centricity affects the types of bottlenecks that arise. At training time, the bottlenecks are generally data-related. Embedded computing relies on scarce sensor data modalities, unlike those commonly addressed in mainstream AI, necessitating solutions for efficient learning from scarce sensor data. At inference time, the bottlenecks are resource-related, calling for improved resource economy and novel scheduling policies. Further ahead, the convergence of AI around large language models (LLMs) introduces additional model-related challenges in embedded contexts. Second, on the domain expansion side, we argue that community expertise in handling resource bottlenecks is becoming increasingly relevant to a new domain: the cloud environment, driven by AI needs. The paper discusses the novel research directions that arise in the data-centric world of AI, covering data-, resource-, and model-related challenges in embedded systems as well as new opportunities in the cloud domain.
This paper presents our experience using Machine Learning (ML) to predict the Worst-Case Execution Time (WCET) of small code snippets on single-core platforms. We provide a concise overview of our work, highlight key observations made throughout our study, and advocate for further exploration of this topic.
The outstanding performance of deep neural networks and machine learning algorithms is driving widespread adoption of these technologies across various application domains, including safety-critical systems like self-driving cars, autonomous robots, and medical diagnostic support systems. However, most deep learning models were not designed to guarantee safe, secure, and predictable behavior. Hence, several challenges must be addressed at multiple architectural levels to ensure their reliability and trustworthiness. This paper discusses some key issues related to AI-powered embedded systems, proposing potential solutions and research directions aimed at enhancing their security, safety, and predictability.
Artificial intelligence (AI) is becoming increasingly relevant in many contexts. In embedded real-time systems, most of the previous research has focused on real-time guarantees for AI workloads (RT-for-AI). Instead, this position paper discusses the potential benefits and application cases of the complementary direction of using AI to optimize real-time systems themselves (AI-for-RT). It presents a vision where AI techniques, such as supervised and reinforcement learning, support system design and online configuration activities that are traditionally addressed using Mixed-Integer Linear Programming (MILP) or heuristic methods. The paper discusses scenarios where AI can potentially outperform classical techniques—such as recursive real-time analysis, systems with complex hardware/software interactions, and dynamic resource management—highlighting the promise of AI in both design-time and runtime real-time systems optimization. Solutions are left to future work: the goal is to populate the “Roadmap Towards Learning-Enabled and Learning-Assisted Real-Time Systems”, which is the target of this special issue.
Real-time scheduling ensures predictability in computing systems, ensuring that tasks meet stringent timing constraints. The integration of machine learning and neural networks into real-time scheduling offers new paradigms for solving constrained optimization problems. This paper briefly explores the intersection of machine learning and real-time scheduling, covering the role of neurodynamic systems, reinforcement learning, and the application of real-time constraints to machine learning models. Additionally, the study discusses challenges in implementing real-time machine learning, including system architectures, accelerators, and safety concerns. The paper concludes with insights into our recent advancements in wearable healthcare systems and secure neural networks in real-time environments.
Integrating AI into real-world systems such as autonomous vehicles or interactive assistants requires the use of compute accelerators. Traditional processors such as x86 or ARM CPUs are insufficient. Unfortunately, real-world systems have responsiveness requirements, and research is underdeveloped on guaranteeing such responsiveness for accelerator-using systems. One constraint has been uncertainty about what sort of accelerator is best for such systems. In this paper, we argue that researchers should focus on the GPU as the accelerator of choice for embedded real-time AI workloads. We argue that GPUs are already being widely adopted, provide leading compute density, and are architecturally well-suited for real-world, real-time systems.
In cyber-physical systems (CPS), it is imperative that tasks are completed within hard deadlines and with some degree of accuracy. Traditional real-time scheduling techniques are concerned with meeting deadlines, but do not concern themselves with the functional performance of tasks. However, when tasks are comprised of machine learning (ML) models, their performance is dependent on the system or environmental state, which can be inferred through the results of other ML tasks in the task set. Additionally, with ML tasks, a functional goal can be achieved by selecting a model among several candidates with different non-functional characteristics. Dynamically selecting the best model to execute given a state estimate may therefore affect the schedulability of other system tasks. We define a task whose functional performance is correlated with the results of another task as logically dependent and propose scheduling task sets composed of ML models with these dependencies in mind to maximize functional performance. We present several examples to show the potential usefulness in CPS and tie this concept with existing literature on real-time scheduling.
