Over the past few years, Federated Learning (FL) has emerged as a prominent paradigm for distributed model training, enabling collaborative learning across heterogeneous edge environments while inherently supporting privacy preservation. Although several frameworks have emerged to support FL workflows, deployment remains a complex and technically demanding phase, often requiring advanced expertise in system administration, virtualization, and networking. This complexity limits accessibility and slows down experimentation and adoption in broader research and industry contexts. To address these limitations, this work introduces a novel deployment toolkit designed to streamline the setup and execution of FL scenarios. Unlike existing solutions, our tool prioritizes ease of use, intuitive configuration, and a gentle learning curve, enabling users to design and launch sophisticated simulations and network topologies without requiring deep system-level knowledge. Through its modular architecture and lightweight orchestration model, the toolkit supports scalable deployments, complex federation strategies, and hybrid edge-cloud environments with minimal overhead.
The progressive digitization of Operation Technology (OT), thanks also to the integration of Industrial Internet of Things (IIoT) devices, is reshaping the landscape of modern industrial systems. This introduces additional complexity and amplifies the heterogeneity across hardware and software ecosystems, thus exacerbating the challenges of orchestration and lifecycle management, which become particularly intricate and demanding. To meet these challenges, modern DevOps practices, such as Continuous Integration and Continuous Deployment (CI/CD), runtime observability, and finegrained access control, are essential to shorten time-to-market, ensure service quality, manage operational complexity, while providing security guarantees. Moving a step toward seamless integration of OT into cloud-native ecosystems, we propose Kubernetes-based OT Orchestrator (KOTO), an orchestration platform that extends Kubernetes to enable comprehensive lifecycle management of OT devices in industrial environments. KOTO bridges the gap between IT and OT by seamlessly integrating with existing CI/CD pipelines and implementing Role-Based Access Control (RBAC) mechanisms tailored to device-level permissions. It abstracts the management of heterogeneous, multi-vendor hardware through a unified abstraction, enabling scalable and secure operations across diverse industrial setups. Without loss of generality, we demonstrate the use of the platform in managing Programmable Logic Controllers (PLCs). We then perform extensive evaluations in a realistic deployment environment, validating KOTO's effectiveness in enhancing operational resilience and responsiveness.
The paper presents a framework for bilingual terminology extraction targeting highly technical domains and the low-resource Kazakh language. We develop two unsupervised term annotators and introduce a hybrid Semantic NMF model that combines non-negative matrix factorization with multilingual embeddings (LaBSE, E5-base, BGE-M3). We further release Matcha, a BIO-annotated English and Kazakh corpus (similar to 3k terms) to fine-tune supervised extractors (BERT-BiLSTM-CRF, XLM-RoBERTa). Across ACTER and ACL RD-TEC 2.0 benchmarks and the new Matcha corpus, UA2 attains up to 51.8% F1, approaching supervised performance; supervised models reach about 70% F1 in Materials Science. The results highlight strong cross-lingual/cross-domain transfer for Semantic-NMF-based unsupervised methods and demonstrate a scalable path to terminology mining and downstream knowledge graph construction in under-represented languages.
Efficient workload forecasting is a key enabler of modern AIOps (Artificial Intelligence for IT Operations), supporting proactive and autonomous resource management across the computing continuum, from edge environments to large-scale cloud infrastructures. In this paper, we propose a Temporal Transformer architecture for CPU utilization prediction, designed to capture both short-term fluctuations and long-range temporal dependencies inworkload dynamics. The model is first pretrained on a large-scale Microsoft Azure VM dataset and subsequently fine-tuned on the Alibaba container dataset, enabling effective transfer learning across heterogeneous virtualization environments. Experimental results demonstrate that the proposed approach achieves high predictive accuracy while maintaining a compact model size and inference times compatible with real-time operation. Qualitative analyses further highlight the model's ability to reproduce workload patterns with high fidelity. These findings indicate that the proposed Temporal Transformer constitutes a lightweight and accurate forecasting component for next-generation AIOps pipelines, suitable for deployment across both cloud and edge intelligence scenarios.
Federated learning (FL) systems are vulnerable to client drift, where evolving client data distributions degrade global model performance. This work introduces DRAF, a drift-aware aggregation strategy that requires no access to client data and addresses both data drift and concept drift. DRAF employs two complementary metrics, a Population Stability Index (PSI)-based score for data drift and a prediction disagreement score for concept drift, combined into a dynamic regularization term to adjust client contributions during aggregation. Experiments on AWS infrastructure using the Flower framework and the CIFAR-10 dataset under controlled drift show that DRAF outperforms standard Federated Averaging (FedAvg), achieving up to a 0.5 reduction in global loss and an average 3% increase in accuracy, demonstrating improved robustness and convergence in non-stationary federated settings.
This paper evaluates the impact of dataset characteristics on the accuracy of temperature prediction models for coupled GPUs in Data Center (DC). Using synthetic data generated via ODEs, we compare eight data generation methods and demonstrate the data content influence prediction accuracy.
Microservices provide modularity and elastic scaling for cloudnative systems, yet highly dynamic workloads make autoscaling both critical and difficult. Bursty traffic surges risk SLA violations, while over-provisioning during low-load periods wastes up to $6.6B annually. An effective autoscaler must not only anticipate workload shifts but also coordinate scaling across interdependent services. Existing ML-, RL-, and GNN-based methods improve efficiency but fail to address the Degradation of Resource Oscillation (DRO)-unstable, repeated up/down scaling triggered by abrupt workload changes or prediction errors-which destabilizes performance and inflates costs. We present GraphPilot, the first Temporal Graph Actor-Critic (TGAC) autoscaler for kubernetes that directly links temporal dependency graph modeling to mitigating oscillations. By representing microservice topologies as evolving graphs and capturing their temporal dependencies through a Temporal Graph Network (TGN), GraphPilot can anticipate LongTerm Dependency Propagation (LTDP)-the underlying cause of DRO-where delayed cross-service dependencies amplify workload shifts and destabilize systems. These temporal graph embeddings are fed into TGAC, enabling coordinated and proactive scaling decisions that stabilize resource allocation under volatile workloads. Extensive experiments on realistic microservice benchmarks show that GraphPilot outperforms state-of-the-art autoscalers in 90.63% of metrics, delivering up to 95.02% higher request-per-pod time (RPT), 136.83% greater throughput, and 73.44% lower latency, while maintaining stable and cost-efficient scaling. The source code is available at https://github.com/ntu-ben/GRLScaler.git.
Modern distributed applications rely on virtualized infrastructures to elastically meet their performance requirements. In this setting, autoscaling enables elastic adaptations at runtime. While allowing for the overcoming of the burden of adjusting the provisioned resources, autoscaling shifts the problem to the definition of an accurate and appropriate threshold, for example, a certain CPU usage, which is a difficult challenge to achieve. Furthermore, defining a priori a fixed value clashes with the dynamicity of modern applications and infrastructure, leading to inflexibility that can affect the quality of service over time. Finally, most autoscaling techniques rely on low, resource-level metrics, which, in complex scenarios, are difficult to gauge. In our paper, we propose HPAQT, a lightweight, stable, and reproducible RL mechanism that self-calibrates the autoscaling threshold to enforce composite, high-level objectives rather than fixed low-level metrics. HPAQT yields an easily interpretable, deployable, and effective policy. In experiments, HPAQT achieves 10x fewer violations than the reference Q-Threshold and beats the standard Kubernetes HPA, with less than 0.5% total violations in over 12 hours, thus demonstrating practical gains.
Machine Learning (ML) services require both responsiveness and efficiency. In the dynamic Computing Continuum (CC), service migration has emerged as an important strategy to optimize service performance, improve power efficiency, or reduce operational costs. A basis enabler of the service migration in the CC is containerization and container orchestration, however, dynamically moving the service among Cloud and Edge servers based on several uncertainties is still challenging. In this paper, we present GreenWise, a framework for intelligent service migration for containerized ML services in the heterogeneous CC. GreenWise extends the monitoring agents to continuously monitor power consumption and performance metrics across diverse layers and sources in real-time, providing a holistic view of states. Leveraging Reinforcement Learning (RL), it enables Power-Performance aware strategies to achieve near-optimal online migration decisions under dynamic conditions. We perform GreenWise on top of a Kubernetes-based CC platform, implementing agents for intelligent migration for containerized ML services. Experimental results demonstrate that GreenWise achieves effective trade-offs between performance and power efficiency. Our proposed Power-Latency-Aware (MaskPPO) migration strategy outperforms Random and Round-Robin baselines 159.2% and 13.8% in a real cluster. These results highlight GreenWise's potential for sustainable and intelligent ML service migration in the dynamic CC.
This paper investigates Knative, an open source serverless framework, and its performance and power consumption behaviours when running different types of applications. As such, it addresses the following research question to gain an understanding of the power consumption and performance benefits Knative may bring to a cloud environment under varying workload conditions: How do different types of serverless applications impact power consumption and performance in Knative Serving, compared to equivalent applications deployed on containers? The experimental results reveal that Knative can eliminate the idle power consumption of deployed applications due to its scale-to-zero capabilities. More notably, for applications operating near CPU and memory limits, using Knative to scale additional instances of the application can reduce response times without proportionally increasing overall power consumption.
Human activity recognition (HAR) in real-world environments is challenging due to sensor noise, environmental variations, and subject-specific differences. While CNNs and RNNs perform well in controlled settings, they often fail to generalize due to limited contextual understanding. We propose a hybrid graph convolutional network (GCN) framework that combines CNN and GCN to enhance context learning and robustness, leveraging GCNs to capture activity dependencies while improving feature extraction. Across controlled and real-world datasets, the CNN-GCN hybrid outperforms baseline CNN and GCN models, achieving higher accuracy and more efficient training in challenging conditions. This underscores the value of context-aware learning in GCN-based HAR for real-world applications.
The IoT-Edge-Cloud computing continuum provides the resources for modern digital services, including, more recently, generative AI applications. These infrastructures, however, have a growing environmental footprint, contributing to CO2 emissions, water use, and resource depletion. As generative AI adoption accelerates, the energy, water, and rare earth demands of training and usage phases are projected to rise significantly in the next years. This poster presents initial results from our efforts to carry out modelling generative AI workloads to create open datasets on resource (memory and CPU) and energy use.
Modern Edge-to-Cloud (E2C) systems require fine-grained observability to ensure adaptive behavior and compliance with performance objectives across heterogeneous and dynamic environments. This work introduces an application-level observability framework that integrates developer-driven instrumentation and SLO-aware feedback for autonomous adaptation. By combining OpenTelemetry, Prometheus, K3s, and Chaos Mesh, the framework enables real-time monitoring and adaptive control across the continuum. A video processing use case demonstrates how application-level metrics guide automatic adjustments to maintain target frame rate, latency, and detection accuracy under variable workloads and injected faults. Preliminary results highlight improved scalability, fault tolerance, and responsiveness, providing a practical foundation for adaptive, SLO-compliant E2C applications.
To prevent insiders from eavesdropping on sensitive information in virtual machines (VMs), recent clouds provide confidential VMs, whose memory is transparently encrypted. Since even confidential VMs cannot protect data from intruders inside them, it is still necessary to use intrusion detection systems (IDS). IDS offloading is used to run host-based IDS outside VMs and prevent IDS from being disabled by intruders. However, offloaded IDS cannot monitor information in the memory of confidential VMs due to memory encryption. This paper proposes SEVmonitor for enabling IDS offloading by running agents inside confidential VMs. Offloaded IDS running in another confidential VM securely obtains memory data from the agent in the target VM. To enhance the security of the agent, SEVmonitor confines a target system in an isolated execution environment created in the target VM and runs the agent outside it. It supports two types of isolated execution environments, a container and an inner VM, to take various tradeoffs. We have implemented SEVmonitor using KVM, Linux, BitVisor, and Xen, and examined monitoring and system performance.
Smart city platforms often use cloud-based infrastructures that leverage containers and virtual machines to manage dynamic demands from diverse data sources, urban services, and users. However, traditional auto-scaling mechanisms in cloud environments typically focus on infrastructure-level metrics, which can result in delayed responses to performance bottlenecks. This paper presents a reactive auto-scaling approach based on the MAPE-K autonomic control loop that integrates infrastructure-level data with applicationlevel metrics. Our approach enables the earlier identification of hotspots in the platform's internal logic, which are often overlooked by traditional techniques. Our solution collects multi-level metrics in a non-intrusive manner to determine whether to scale resources vertically or horizontally before significant resource contention arises. We validated our approach in a real-world smart city platform that manages geospatial data and experiences highly variable workloads. Experimental results show that our solution can reduce reaction times and prevent unnecessary scaling actions.
Federated Learning (FL) enables model training across decentralized datasets, preventing the centralization of sensitive raw data. Despite robust algorithmic progress, operational adoption is often hindered by technological heterogeneity in collaborative environments, where different teams employ diverse Machine Learning (ML) frameworks (e.g., PyTorch, TensorFlow, JAX), whose differences in internal model representation hinder direct collaboration. This paper presents an empirical analysis of FL-as-API, a ServiceOriented Architecture (SOA) for FL designed to operate in these heterogeneous environments. The system enables inter-framework operability via an orchestrator that performs real-time weight normalization, assuming the responsibility for "translating" models between different formats. The architecture's performance is empirically evaluated on a purpose-built FL environment featuring clients implemented in both PyTorch and TensorFlow, demonstrating that: 1) overall system performance (measured in rounds per minute) is primarily constrained by the computational time of the slowest client, rather than by communication or aggregation; 2) the cost of interoperability (server-side weight normalization) is quantitatively low compared to the total round duration; and 3) the service-level architecture enables the effective identification of client stragglers, isolating performance bottlenecks at the individual machine level. This work provides a quantitative evaluation of the costs and benefits of architectural flexibility in real-world FL systems.
Traditional static computational resource allocation in cloud or on-premises clusters often results in inefficient overprovisioning. Users frequently lack precise knowledge of the memory and processors their applications require, leading them to request excess resources. This causes wasted capacity, higher costs, and, in shared environments, longer queue waiting times. Dynamic resource allocation through autoscaling addresses this issue by adjusting resources at runtime. Kubernetes, a widely used container orchestration platform, supports autoscaling via Horizontal and Vertical Pod Autoscalers. However, its default restart-based scaling can disrupt stateful, long-running workloads without checkpointing. This work leverages Kubernetes' new in-place scaling, which resizes resources without restarts, to propose the Dynamic Resizing Strategy (DRS), a novel autoscaling approach that proactively manages contention by temporarily throttling co-located pods to prioritize a bursting application. We evaluate it with NAS Parallel Benchmarks and synthetic workloads in co-execution scenarios, showing improved efficiency and stability, increasing success rates and reducing global average wait time by over 18% compared to the Burstable QoS class.
Cloud computing has transformed application deployment and scalability, with Kubernetes emerging as the dominant platform for container orchestration. However, optimizing CPU resource allocation during container startup remains a challenge, particularly for applications utilizing the Java Virtual Machine (JVM), which exhibit high CPU demands during initialization. Insufficient CPU provisioning can cause significant startup delays and performance throttling, while excessive allocation inflates cloud costs, underscoring the necessity for optimal resource management. This paper introduces a predictive startup CPU resource allocation framework integrated with Google's Kube Startup CPU Boost controller, leveraging Kubernetes' In-place Resource Resize feature to dynamically allocate CPU resources at container startup. Inspired by IBM's MAPE-K (Monitor, Analyze, Plan, Execute, Knowledge) framework, our proposed system CLUEAI, analyzes images, monitors container startup behavior and predicts optimal CPU Boost levels and durations. It employs a hybrid approach combining machine learning and statistical methods. A machine learning model is used for initial predictions of startup CPU resource requirements based on image metadata, eliminating the need to run the application beforehand. Statistical methods further refine these predictions in subsequent container startups, making the system self-adaptive. To train and validate the predictive model, a comprehensive dataset comprising container metadata and runtime performance metrics was developed. Based on experimental results, the CLUEAI system reduced the application startup time by up to 40%, with an average decrease of 19.9% and a reduction in CPU wastage of 31.74% compared to un-boosted deployment scenarios.
Compute clusters are major power consumers in Cloud and Edge data centers, making it critical to reduce power usage and costs without compromising service levels objectives. Energy Performance Preference (EPP) settings and CPU frequency scaling can lower power but typically at the cost of reduced performance. When considering clusters with heterogeneous power profiles, it is essential to map workloads to the most suitable profile based on their quality-of-service constraints. Current orchestrators overlook power-profile heterogeneity; this is a particular concern at the Edge, where otherwise identical hardware may range from poweroptimized to performance-oriented yet remain indistinguishable to schedulers. We present a taxonomy of power-aware orchestration, and extend the default Kubernetes scheduler with power-profile awareness. We evaluate the feasibility of this extended scheduler by comparing three power profile-aware scheduling strategies on a testbed running a microservices benchmark, with results showing that average power use can be reduced by up to 12% while maintaining application performance. We conclude with key challenges and future research directions.
LoRaWAN Low Power Wide Area Networks have become a standard for remote sensing over large geographical areas with limited access to energy; however, this low-power capability inherently limits communication throughput, posing scalability challenges as network density increases. Edge intelligence (edge AI) can reduce network traffic by deploying lightweight models locally that classify and filter local data before transmission to remote servers. However, models trained with limited local data that run on devices with restricted memory and execution capability necessitate continuous updates to remain accurate within dynamic environments. Edge intelligence methods, such as federated learning, transfer global model updates to end devices for incremental training and local adaptation. While model transmission is possible using Firmware Update Over The Air (FUOTA) in LoRaWAN, current implementations necessitate manual selection of transmission parameters that highly impact update duration and energy efficiency. This paper proposes a novel approach for updating models over LoRaWAN, addressing distinct and significant challenges compared to typical scenarios over Wi-Fi or cellular networks. First, we fully automate the FUOTA process to propagate model updates to selected machines with heterogeneous communication capabilities. Then, we propose three parameter selection policies to balance the energy consumed by devices and the time taken to update entire networks. An evaluation on a testbed demonstrates the effectiveness of our approach in propagating updates automatically. Large-scale simulations of up to 300 end devices show that our energy-oriented update policy reduces the energy consumption of end devices between 2.2x and 2.7x compared to state-of-the-art baselines.
