Debugging high-dimensional state spaces in cloud continuum environments poses significant challenges, particularly when investigating extreme conditions such as high latency, competing on resources, or configuration anomalies. This paper presents a novel supervised machine learning-based approach to efficiently assist the debugging process by steering toward potential fault states in an automated way. Leveraging typical blueprint primitives, such as load balancers and temporal data storage in the presented case studies, Multi-Layer Perceptron (MLP) and Dense Neural Networks (DNN) were trained to predict the distance to extreme situations. The trained model informs a traversal mechanism that explores the state space using this heuristic, minimizing the time and consumed resources required to detect actual faults. The first experiments conducted with two foundational blueprint primitives (buffers and multi-tier load balancers) demonstrate the promising effectiveness of the approach in locating potential fault states. By integrating this method into cloud-edge debugging tools, developers can enhance not only fault localization but reliability and performance as well, particularly for extreme timing conditions. Future work will explore a wider set of primitives, as well as adjacency matrix representations and convolutional techniques, to improve applicability, scalability and robustness of the presented solution.

Nowadays, the number of server farms is exploding in many parts of the world. This is damaging the environment since it leads to significant carbon emissions, requires a lot of water for cooling, increases electromagnetic pollution and leads to the destruction of more and more natural habitats. This growth cannot be stopped and with the spread of artificial intelligence, many more server resources will be needed around the world. Our research aims to help optimise these resources. The distribution of applications between servers usually happen manually, but various orchestration solutions (Docker Swarm, Kubernetes), already provide this step automatically. Our research investigates their distribution efficiency and compares it with our own solution to estimate the resource requirements of open source softwares using machine learning. Since the mentioned orchestration tools do not perform this kind of estimation, our research provides new results in this area.

In conventional supervised learning of neural networks, training samples are selected either randomly or in a predefined order, assuming independence across samples. This paper diverges from that setting by embedding the learning process within a dynamic control system. Specifically, we consider a discrete-time control system where the output is given by a nonlinear mapping, that dynamically adjusts the number of virtual machines (VMs) based on workload characteristics such as CPU, memory, and network usage. The system’s output is determined by a neural network that estimates the deviation from a target utilization profile. In online supervised learning embedded in feedback control, data generation is shaped by model performance, leading to a narrowing of the observed input distribution over time. This self-induced sampling bias may reduce model robustness, stability and adaptability. We demonstrate that simple periodic perturbations to the VM allocation process act as an effective form of regularization, improving learning robustness without relying on external reward or replay mechanisms. Unlike traditional approaches using fixed training sets, in our formulation the system operates online where at each time step, multiple candidate control inputs u[k] ∈ U are evaluated continuously and each yielding a predicted output y[k] = f(x[k]). At each step, the controller selects the action that minimizes the predicted deviation from the desired reference, which then determines the next state x[k + 1] and yields the next training sample for the neural network. As a result the learning trajectory is not predetermined but is dynamically created by the controller’s actions, which depend on the network’s current predictions. We present how this closed-loop interaction between prediction and sample selection influences learning stability, convergence, and input space coverage in an online setting.

Debugging cloud services is increasingly challenging due to their distributed, dynamic, and scalable nature. Traditional methods struggle to handle large state spaces and the complex interactions between microservices, making it difficult to diagnose failures and identify critical components. This paper presents a Graph Neural Network (GNN)-based approach that enhances cloud service debugging by predicting system-level fault probabilities and providing interpretable insights into failure propagation. Our method models microservice interactions as graphs, where failures propagate probabilistically. Using Markov Decision Processes (MDPs), we simulate failure behaviors, capturing the probabilistic dependencies that influence system reliability. The trained GNN not only predicts fault probabilities but also identifies the most failure-prone microservices and explains their impact. We evaluate our approach on various service mesh structures, including feature-enriched, tree-structured, and general directed acyclic graph (DAG) architectures. Results indicate that our method is effective in the operational phase of cloud services, enabling proactive debugging and targeted optimization. This work represents a step toward more interpretable, reliable, and maintainable cloud infrastructures.

Nowadays more and more servers are put to use due to the explosive improvement of information technology all around the world. Smart solutions, big data storages and cloud technologies brings with it the ever increasing energy consumption. It is often heard on the news that the energy consumption of some newly built server farm is getting to the level of a smaller city. The question of the near future is the efficient use of energy and avoiding wasteful usage of it. In the context of softwares this can be achieved by running the softwares in optimized hardware environments, using the capacity of the serves to the maximum, allocating more procedures to them until the point where it does not affect performance. In this study we will build a training pattern from hundreds of program codes, vectorizing them based on their main characteristics. Under a dynamic stress test their ideal resource consumption will be determined, then a deep neural network will be trained to determine the minimal resource allocations needed ( memory size, cpu cores) to achieve optimal runtime for a program code that is unknown to it. Based on this it would be possible to develop an optimal distributing algorithm that would place the services on a whole server farm in a way to achieve the desired quality of service using the least amount of hardware.

Performance and cost-effectiveness are sustained by efficient management of resources in cloud computing. Current autoscaling approaches, when trying to balance between the consumption of resources and QoS requirements, usually fall short and end up being inefficient and leading to service disruptions. The existing literature has primarily focuses on static metrics and/or proactive scaling approaches which do not align with dynamically changing tasks, jobs or service calls. The key concept of our approach is the use of statistical analysis to select the most relevant metrics for the specific application being scaled. We demonstrated that different applications require different metrics to accurately estimate the necessary resources, highlighting that what is critical for an application may not be for the other. The proper metrics selection for control mechanism which regulates the requried recources of application are described in this study. Introduced selection mechanism enables us to improve previously designed autoscaler by allowing them to react more quickly to sudden load changes, use fewer resources, and maintain more stable service QoS due to the more accurate machine learning models. We compared our method with previous approaches through a carefully designed series of experiments, and the results showed that this approach brings significant improvements, such as reducing QoS violations by up to 80

With the increasing utilisation of machine-learning algorithms, autoscaling methods have the potential to provide more sophisticated control mechanisms for the cloud application operators over virtualized resources in terms of provisioning. This paper introduces machine learning-based autoscaling approach introducing several improvements including metric selection, proactivity, various enhancements of the neural network among others. Our enhanced machine learning models enables the autoscaler algorithm to react more quickly to sudden load changes, increase proactivity while using fewer resources and reducing Quality of Service (QoS) violations for cloud-based services. The comprehensive measurements indicate that QoS violations may reduce by up to 80%, while the level of resource utilization either remains constant or decreases slightly (by 3-4%) in certain applications. In cases where there was no reduction in QoS violations, the utilization of resources saw a significant decline, falling between 20-50%. The proposed changes have been analyzed and tested under various conditions, representing 3 distinct and common use cases in cloud environments. This developed process has been implemented on the science cloud of the Hungarian Research Network supporting the operation of the infrastucture that hosts over 300 scientific projects.

Orchestration methods at Infrastructure-as-a-Service (IaaS) level automate the deployment, scaling, and management of virtualized resources, typically across multiple hosts and data centres. While orchestration provides many advantages, itealso introduces several challenges in testing and debugging phases, particularly due to the distributed nature of the virtualized resources. Even the proper initial deployment of interdependent virtual machines (VM) may cause fatal errors since the unpredictable timing conditions may change the overall initialisation method, which can lead to abnormal behaviour, i.e. inecomplex, non-deterministic environments, the set of VM configurations can drift from their expected states (‘configuration drift’). The overall motivation of our research is to improve the reliability of cloud-based infrastructures with minimal user interactions and significantly automate the time-consuming debugging process. This paper focuses on the examination and behaviour of cloud-based infrastructures during their deployment phase. Weecontinued the adaption of a replay-active control based debugging technique, called macrostep, inethe field of cloud orchestration. Ineorder to provide efficient support for developers troubleshooting major deployment related errors, the fundamental macrostep mechanisms have been enriched and significantly extended including 1) the automated generation of collective breakpoint sets, 2) parallel and robust traversal method for such consistent global states with 3) automated evaluation of global predicates in each global state of VM set. Furthermore, the novel methods have been 4) generalized towards wider user scenarios by targeting the Terraform orchestration tool as well (besides the already supported Occopus). The paper describes the significantly enhanced approach, our design choices, and also the implementation of the experimental debugger tool with a use case for validation purposes by addressing the deployment of a SLURM (HPC) cluster.

Collecting and analysing large amounts of data in the Cloud-to-Edge computing continuum raises novel challenges that traditional centralised orchestration solutions cannot handle efficiently. To overcome the limitations of current centralised application management approaches, this paper presents a fully decentralised application-level orchestrator, based on the notion of self-organised interdependent Swarms. Application microservices are managed in a dynamic Orchestration Space by decentralised Orchestration Agents, governed by distributed intelligence that provides matchmaking between application requirements and resources, and supports the dynamic self-organisation of Swarms. Knowledge and trust, essential for the operation of the Orchestration Space, are managed through blockchain-based trusted solutions and the utilisation of emerging methods such as Self-Sovereign Identities (SSI) and Distributed Identifiers (DID). End-to-end security of the overall system is assured by utilising state-of-the-art cryptographic and privacy-preserving data analytics algorithms. A digital twin, that runs in parallel to the physical system, further improves its behaviour with predictive feedback. The presented concept is going to be implemented in the EU-funded Swarmchestrate project that starts in 2024.

The manual limitation of software-related resources (memory quantity, processing time) might be challenging even in modern desktop operating systems or on server computers. The applications are always competing for the available resources and those of greater need for resources may starve the others. In case of server computers, outages can happen in this period. A previous solution for this problem was to install resource intensive applications on a physically separate computer therefore they did not affect other applications. With the appearance of virtualisation on one physical computer more virtual machines are running simultaneously. Each of their resources can be limited separately. In this case the abundance of virtual machines means the unnecessary waste of resources. Another more suitable method is the use of containers which run on the operating system’s kernel, however these provide an isolated, limitable environment. The addressed problem to solve is the estimation of resources needed which even nowadays is often working on an empirical basis. In our research we are analysing the source code of applications by static code analysis technique, making a composite call graph and a dependency graph. With the descriptors and characteristics of the graphs there is an opportunity to train neural networks the programmed estimation of required resources. [1]

The emerging availability of quantum compute resources fosters the examination of its exploitation possibilities in different scientific domains, like artificial intelligence, manufacturing or finance. A significant number of research scientists primarily rely on cloud computing infrastructures while conducting research, whereas access to real (mostly remote) quantum hardware resources requires the deployment and proper configuration of different software components. In this paper we present a hybrid, cloud-based reference architecture that lowers the entry barrier to start new experiments with a wide range of quantum compute resources. The solution facilitates the execution of distributed quantum computing simulations in traditional cloud environments, and also access to several remote quantum compute resources. The reference architecture is highly portable to various cloud platforms, resulting in efficient adaptation and application possibilities by research communities. The paper describes our related experiences using commercial cloud providers and on the federated, OpenStack-based research infrastructure of the Hungarian Research Network (abbreviated as HUN-REN).

Fault detection, which requires a lot of time and complexity, is one of the most difficult tasks for cloud computing. In this research, we investigate the utilization of the anytime algorithm in cloud-based fault detection techniques. Anytime algorithms are real-time decision-making algorithms that focus on time-deliberation. In this paper, the basic concepts of the Anytime algorithm in existing works as well as its benefits of usage are analyzed. Various literature is reviewed and combined for possible fault recognition and time deliberation. The paper includes details of anytime processing in fault detection in different subject areas. Based on the survey, we aim to define the algorithm processing implementation as future work.

One effective method for assessing the dependability of computer systems is fault injection. This deliberate technique introduces faults into a system to assess its resilience and ability to handle abnormal conditions. Therefore, this study investigates and simulates the different network problems in the CloudSim Plus environment. CloudSim Plus is a simulation framework that enables the modeling and simulation of cloud computing environments, allowing researchers and practitioners to evaluate the performance and behavior of cloud-based systems and algorithms. Network fault detection and its management are essential duties in cloud systems. Moreover, the feasibility of manual monitoring and involvement has decreased as these infrastructures expand and change. This paper briefly introduces network problems and fault injection outcomes in CloudSim Plus nodes.

Cloud architecture blueprints or reference architectures allow the reuse of existing knowledge and best practices when creating new cloud native solutions. Therefore, debugging of reference architecture candidates (or their new versions) is an extremely crucial but tedious and time-consuming task due to the deployment of complex services in typical multi-tenant and non-deterministic environments. During the debugging/testing/maintenance scenarios, we might be able to achieve greater levels of test coverage (and eventually improved reliability) by modelling and verifying at least their most fundamental building blocks and their interconnections. The main objective of our work is to integrate stochastic modelling and verification techniques based on deep learning methods into the debugging cycle in order to handle large state spaces more efficiently, i.e. by steering the process of traversing state space towards suspicious situations that may result in potential bugs in the actual system with smart steering during the traversal. For this purpose, our presented and illustrated approach combines (among others) Continuous Time Markov Chain modelling (CTMC) techniques with deep learning methods including autoencoder, Long Short-Term Memory (LSTM) and Graph Neural Network (GNN) models. Our experiences are summarized with widespread cloud design patterns including load balancing and service mesh topologies. According to the results, the debugging cycle can be partly automated through the application of deep learning methods. The autoencoders are able to detect erroneous load balancer behaviors (anomalies) in complex configurations; the LSTMs demonstrate implicitly some random nature of the inspected processes, and GNNs exploit the additional topology-related information in service meshes.

This work introduces a scalable, cloud-agnostic and fault-tolerant data analytics platform for state-of-the-art autonomous systems that is built from open-source, reusable building blocks. As the baseline for further new reference architectures, it represents an architecture blueprint for processing, enriching and analyzing various feeds of structured and non-structured input data from advanced Internet-of-Things (IoT) based use cases. The platform builds on industry best practices, leverages on solid open-source components in a reusable fashion, and is based on our experience gathered from numerous IoT and Big Data research projects. The platform is currently used in the framework of the National Laboratory for Autonomous Systems in Hungary (abbreviated as ARNL). The platform is demonstrated through selected use cases from ARNL including the areas of smart/autonomous production systems (collaborative robotic assembly) and autonomous vehicles (mobile robots with smart vehicle control). Finally, we validate the platform through the evaluation of its streaming ingestion capabilities.

This review article summarizes the history the Hungarian Scientific Cloud Infra-structure project. This research infrastructure was launched officially on 1 October 2016, funded by the Hungarian Academy of Sciences. With the support of ELKH, the infrastructure's capacity has been substantially boosted; the features and workflows that it offers to scientists were significantly expanded to celebrate the arrival of the year 2022. The article reviews the types of work Hungarian researchers implemented on the infrastructure, thereby providing an overview of the state of cloud-computing enabled science in Hungary.

Safe motion together with improved economy and traveling performance levels are important requirements against automated vehicles. Thus, the design of enhanced control systems is requested, which contain conventional model-based controllers and the use of unconventional approaches, e.g., learning features and cloud-based methods. This paper proposes a hierarchical vehicle control design method with learning functions, which incorporates control in two levels, such as in cloud level and in vehicle level. The control on the cloud level is designed by using reinforcement learning, with which the maximum speed for the vehicle is achieved. The vehicle level contains a robust controller and a supervisor, with which the collision avoidance of the vehicle is guaranteed. The hierarchical control guarantees performance requirement of safe motion, i.e., collision avoidance in all scenarios, even if the connection with the cloud is lost. The proposed control on indoor Hardware-in-the-Loop platform is implemented. The effectiveness of the control and the safe motion of the vehicle under various scenarios with and without cloud connection are demonstrated.

Fault detection and handling are crucial tasks in cloud systems. As these infrastructures are growing and evolving, manual monitoring and interaction have become less feasible. To deal with this issue, monitoring systems are developed to track the behavior of the various components (e.g. nodes) in cloud systems, as well as the served applications in the virtual environment. Nowadays, most cloud environments provide graphics accelerators for their users leading to different problems. However, the application of GPUs in deep learning could also help the detection of incorrect behavior. In this paper, a short overview of cloud monitoring and fault detection methods is given focusing on GPU-enabled nodes.

Flat-field correction (FFC) is commonly used in image signal processing (ISP) to improve the uniformity of image sensor pixels. Image sensor nonuniformity and lens system characteristics have been known to be temperature-dependent. Some machine vision applications, such as visual odometry and single-pixel airborne object tracking, are extremely sensitive to pixel-to-pixel sensitivity variations. Numerous cameras, especially in the fields of infrared imaging and staring cameras, use multiple calibration images to correct for nonuniformities. This paper characterizes the temperature and analog gain dependence of the dark signal nonuniformity (DSNU) and photoresponse nonuniformity (PRNU) of two contemporary global shutter CMOS image sensors for machine vision applications. An optimized hardware architecture is proposed to compensate for nonuniformities, with optional parametric lens shading correction (LSC). Three different performance configurations are outlined for different application areas, costs, and power requirements. For most commercial applications, the correction of LSC suffices. For both DSNU and PRNU, compensation with one or multiple calibration images, captured at different gain and temperature settings are considered. For more demanding applications, the effectiveness, external memory bandwidth, power consumption, implementation, and calibration complexity, as well as the camera manufacturability of different nonuniformity correction approaches were compared.