Edge computing is poised to become a cornerstone of the emerging 6G landscape, where an ever-growing class of ultra-low-latency applications must be served close to the user. Despite its promise, real-world deployments remain nascent, with large-scale implementations anticipated by both Communication and Digital Service Providers (CSPs/DSPs) within the following years. Consequently, strategic edge network design is essential not only to maximize performance, but also to avoid redundant investments that can lead to an increased sum of Capital and Operational Expenditures (CAPEX/OPEX). In this work, we address a tri-fold problem: (i) the selection of deployment locations, (ii) the configuration of devices at the chosen location sites, and (iii) the assignment of the projected workload. Our objective is formulated as a weighted combination of the edge infrastructure’s establishment cost, the expected cumulative workload latency and the total expected energy consumption in the operational phase. To capture the spatial and temporal variability of demand, we solve the assignment subproblem over distinct snapshots, each representing a unique workload projection. We first present a Mixed Integer Linear Programming (MILP) formulation that yields the optimal solution; however, due to its computational intractability, we propose a novel adaptation of the Multi-Objective Evolutionary Algorithm by Decomposition (MOEA/D), with an embedded heuristic algorithm to assist in the chromosome fitness calculation. This method leverages the similarity among neighboring subproblems in a multi-objective framework to efficiently approximate the underlying Pareto frontier. In the experiments, the proposed method is contrasted with a sophisticated single-objective Rollout approach. Our results demonstrate the benefits of adopting a multi-objective algorithm in terms of performance, stability and interpretability across different scalarized subproblems. The proposed framework offers a practical intent-based decision support tool for edge infrastructure providers, weighing CAPEX against operating objectives ahead of initial deployment.

The complexity of modern applications necessitates the decomposition of workloads into logically dependent subtasks. As infrastructures move closer to data sources to decongest backbone networks and reduce communication delays, task deployment and scheduling become increasingly challenging. Orchestrators must respect spatial and temporal dependencies among subtasks and computing nodes, while optimizing goals such as latency and energy consumption. As edge adoption remains limited, operators pursue cooperative solutions that share resources without requiring significant investment. We consider a collaborative infrastructure, split into multiple domains, each with proprietary resources, and a shared pool for all service demands. Multiple agents operate concurrently with partial knowledge of the system, cooperating to allocate shared resources efficiently. In this work, we propose a Multi-Agent Reinforcement Learning scheduler that leverages Graph Neural Networks to handle spatiotemporal dependencies and uses lightweight inter-domain messaging for inter-domain cooperation. The learned policy is scalable and effective, reducing execution time and energy, increasing parallelism and mitigating congestion in simulations with real workload data.

Edge computing brings processing closer to data sources, reducing latency and bandwidth usage for modern applications. However, the limited capacity of edge resources and volatile nature of workload demands create significant challenges for efficient resource management, often leading to resource underutilization. In this work, we propose a speculative resource allocation framework supported by stochastic workload forecasting, inspired by the Black-Scholes financial model. This framework dynamically assesses the risk associated with fluctuating demands over different time windows and proactively aligns resource allocation based on the performed risk assessments. The outcomes of this model drive a multi-objective heuristic mechanism that dynamically manages resources, speculatively aligning differing workload demands when placing them within a node, optimizing key performance metrics such as latency, infrastructure utilization, costeffectiveness, and potential application disruptions during execution. Our approach does not require training, making it more adaptable to fluctuating demands compared to machine learning-based methods. Through simulations we demonstrate that our framework improves performance and resource utilization, providing a scalable, responsive, and cost-effective solution that benefits both end-users and operators.

Edge computing has consolidated as an essential technology for addressing the stringent requirements of modern applications, by distributing computing resources closer to data sources. Nonetheless, this innovation introduces significant challenges for the infrastructure designers and operators, given the high number of edge locations, the heterogeneity of edge resources and the varying requirements of today’s applications. Effective edge-network design is critical to harnessing its full potential, ensuring optimal performance, resource availability and cost efficiency during the applications’ execution. In this work, we propose mechanisms that address the challenge of joint optimal edge deployment location and capacity and device configuration, with respect to workload constraints. We formulate the respective problem as a multi-objective optimization that simultaneously considers the activation and resource costs, energy efficiency, and the workload’s experienced latency. Initially, we present the Mixed Integer Linear Programming (MILP) formulation that yields the optimal solution. To tackle its increased computational complexity, we also propose a rollout mechanism. It iteratively leverages a best-fit heuristic to perform the resource allocation and thus evaluates the impact of different deployment schemes on the overall system performance in a reinforcement learning manner. Our simulation experiments demonstrate the effectiveness of the developed mechanisms in enhancing responsiveness, reducing energy consumption and optimizing the Return On Investment (ROI) of the infrastructure across various deployment scenarios.

Abstract The orchestration of cloud computing infrastructures is challenging, considering the number, heterogeneity and dynamicity of the involved resources, along with the highly distributed nature of the applications that use them for computation and storage. Evidently, the volume of relevant monitoring data can be significant, and the ability to collect, analyze, and act on this data in real time is critical for the infrastructure’s efficient use. In this study, we introduce a novel methodology that adeptly manages the diverse, dynamic, and voluminous nature of cloud resources and the applications that they support. We use knowledge graphs to represent computing and storage resources and illustrate the relationships between them and the applications that utilize them. We then train GraphSAGE to acquire vector-based representations of the infrastructures’ properties, while preserving the structural properties of the graph. These are efficiently provided as input to two unsupervised machine learning algorithms, namely CBLOF and Isolation Forest, for the detection of storage and computing overusage events, where CBLOF demonstrates better performance across all our evaluation metrics. Following the detection of such events, we have also developed appropriate re-optimization mechanisms that ensure the performance of the served applications. Evaluated in a simulated environment, our methods demonstrate a significant advancement in anomaly detection and infrastructure optimization. The results underscore the potential of this closed-loop operation in dynamically adapting to the evolving demands of cloud infrastructures. By integrating data representation and machine learning methods with proactive management strategies, this research contributes substantially to the field of cloud computing, offering a scalable, intelligent solution for modern cloud infrastructures.

Edge computing has emerged as a computing paradigm where the application and data processing takes place close to the end devices. It decreases the distances over which data transfers are made, offering reduced delay and fast speed of action for general data processing and store/retrieve jobs. The benefits of edge computing can also be reaped for distributed computation algorithms, where the cloud also plays an assistive role. In this context, an important challenge is to allocate the required resources at both edge and cloud to carry out the processing of data that are generated over a continuous (“infinite”) time horizon. This is a complex problem due to the variety of requirements (resource needs, accuracy, delay, etc.) that may be posed by each computation algorithm, as well as the heterogeneous resources’ features (e.g., processing, bandwidth). In this work, we develop a solution for serving weakly coupled general distributed algorithms, with emphasis on machine learning algorithms, at the edge and/or the cloud. We present a dual-objective Integer Linear Programming formulation that optimizes monetary cost and computation accuracy. We also introduce efficient heuristics to perform the resource allocation. We examine various distributed ML allocation scenarios using realistic parameters from actual vendors. We quantify trade-offs related to accuracy, performance and cost of edge/cloud bandwidth and processing resources. Our results indicate that among the many parameters of interest, the processing costs seem to play the most important role for the allocation decisions. Finally, we explore interesting interactions between target accuracy, monetary cost and delay.

Cell-Free (CF) technology is considered as a candidate to support the “5G and beyond” networks, mitigating the limitations of the traditional networks in terms of flexibility and intercell interference. These networks consist of distributed Access Points (APs) that form clusters, and co-operate in time to serve the User Equipment (UE) demands. The number of the AP that participate in a cluster and the level at which they cooperate impacts both the achieved spectral efficiency and the size of the utilized communication and processing resources, which in most cases are scarce and limited. In our work, we propose mechanisms that perform joint allocation of fiber and wireless resources in a converged fiber-wireless infrastructure, consisting of a TWDM PON, mMIMO Base Stations and CF. To perform the joint allocation of the wireless and wired resources, we propose an optimal Mixed Integer Linear Program (MILP). As the complexity is high and the execution time prohibitively long for real size scenarios, we also present a multi-agent rollout mechanism to efficiently tradeoff execution time with performance.

We present an end-to-end solution to facilitate the seamless execution of hardware-accelerated compute-intensive tasks on heterogeneous hardware platforms spanning the Cloud-Edge continuum. Our approach includes a programming interface, orchestration, application management components, the vAccel framework, and a library of hardware-accelerated kernels. These components enable a Function-as-a-Service (FaaS) based operational flow that supports numerous diverse use cases while minimizing the time required for the developer to integrate their code and for the vendor to provide hardware acceleration capabilities to end users. Experimental results showcase the merits of our approach.

Distributed storage systems spanning across different cloud data centers have substantially improved availability and flexibility for data storage and retrieval operations. However, stringent latency requirements of emerging applications necessitate optimized selection of storage resources that exhibit smaller delay. Introducing edge resources into distributed storage systems enables data placement closer to its source, but simultaneously increases the complexity of decision-making and orchestration processes for optimal data placement. In this work, we develop mechanisms for storing data across an infrastructure that includes both edge and cloud resources. Our approach focuses on optimizing data integrity, longevity, security, and cost, while leveraging erasure coding when performing the resource allocation. We first present a comprehensive mixed integer linear programming formulation of the storage resource orchestration problem. As the search space for the optimal solution can be vast and the execution time prohibitively large for real size problems, we also propose an innovative multi-agent heuristic approach that uses the rollout, a reinforcement based policy, to balance performance and execution time efficiently. Through various simulation experiments, we evaluate the developed mechanisms and trade-offs involved in our approach. By incorporating data from a multi-cloud provider, we further enhance the validity of the simulations and the conclusions drawn.

Edge computing has emerged as a paradigm for local computing/processing tasks, reducing the distances over which data transfers are made. Thus, an opportunity is presented for data transfer-intensive, distributed machine learning. In this paper we develop a solution for serving distributed Machine Learning (ML) training jobs at the edge– cloud continuum. We model the specific requirements of each ML job, and the features of the edge and cloud resources. Next, we develop an Integer Linear Programming algorithm to perform the resource allocation. We examine different scenarios (different processing and bandwidth costs) and quantify tradeoffs related to performance and cost of edge/cloud bandwidth and processing resources. Our simulations indicate that even though there are many parameters that determine the allocation, the processing costs seem to play on average the most important role. The cloud b/w costs can be significant in certain scenarios. Finally, in certain examined cases, significant monetary benefits can be achieved through the collaboration of both edge and cloud resources when compared to using exclusively edge or cloud resources.

The uncertain and non-dispatchable nature of renewable energy sources renders Demand Response (DR) a critical component of modern electricity distribution systems. Demand Response (DR) service provision takes place via aggregators and special distribution-level markets (e.g., flexibility markets), where small, distributed DR resources, such as building energy management systems, electric vehicle charging stations, micro-generation and storage, connected to the low-voltage distribution grid, offer DR services. In such systems, energy balancing (and thus, also DR decisions) have to be made close to real-time. Thus, market clearing algorithms for DR service provision must fulfill several requirements related to the efficiency of their operation. More specifically, a DR market clearing algorithm needs to be optimal in terms of cost-efficiency, scalable in terms of number of assets and locations, and able to satisfy real-time constraints. In order to cope with these challenges, this article presents a distributed DR market clearing algorithm based on Lagrangian decomposition, combined with an optimal cloud resource allocation algorithm for assigning the required computation power. A heuristic algorithm is also presented, able to achieve a near-optimal solution, within negligible computational time. Simulations, performed on a testbed, demonstrate the computational burden introduced by various DR models, as well as the heuristic algorithm's near-optimal performance. The resource allocation algorithm is able to service multiple DR requests (e.g., in multiple distribution networks), and minimize the cost of computational resources while respecting the execution time constraints of each request. This enables third parties to offer cost-efficient and competitive DR operation as a service.

The design and operation of any network are complex processes that require the evaluation, utilization and configuration of a variety of usually expensive network devices. Through the use of an emulation platform, network operators are able to examine different scenarios and network parameters and benefit from multi-objective decision mechanisms. These enable the decrease of the network design phase duration and the optimal operation of the network under different well examined conditions. In this work, we present an emulation platform for SDN-enabled 5G integrated Fiber-Wireless networks that provides a transparent view of the 5G infrastructure to any SDN-based control plane. We present the overall architecture and design of the emulator, along with the implementation details of its main components. Network devices are described through YANG models and are emulated using containerized processes, configured and managed through the Network Configuration (NETCONF) protocol. Finally, a number of emulation scenarios are described and evaluated, utilizing a joint fiber and wireless resource allocation algorithm that drives the SDN-enabled devices.

The efficient allocation of network resources is key to the overall performance and the quality of services provided. Thanks to their high data rates, optical networks are the cornerstone of present and future core, metro, access, and datacenter networking. Many works in the field, formulate resource allocation operations as offline combinatorial problems, assuming a known static traffic matrix, and use integer linear programming (ILP) as well as heuristics. In contrast, other works assume randomly generated traffic and propose online schemes that serve connection requests one by one. In practice, traffic in optical networks is neither static nor completely random, but is usually semi-periodic, following some (e.g., daily or weekly) pattern. We present a traffic-pattern-driven approach for elastic optical networks for serving immediate and in advance network requests, where the decisions of an offline process, optimizing resource allocation for the traffic pattern expected during an epoch (day, week, etc.), are analyzed and then drive the operation of an online process that serves requests one by one, as they arrive. In this way, the online mechanism's decisions come close to the optimal ones, if the traffic pattern indeed repeats itself to some extent, while its execution time remains small. We present two alternatives of this approach, the exact and the relative, based on the way the offline mechanism's decisions are analyzed and translated to online actions. Our simulation results exhibit the performance benefits of the pattern-driven approach under various traffic conditions.

The monitoring and the analysis of public clouds gains momentum, due to their widespread exploitation by individual users, researchers and companies for their daily tasks. We propose an algorithm for optimizing the cost and the utilization of a set of running Amazon EC2 instances by resizing them appropriately. The algorithm, namely Cost and Utilization Optimization (CUO) algorithm, receives information regarding the current set of instances used (their number, type, utilization) and proposes a new set of instances for serving the same load, so as to minimize cost and maximize utilization, or increase performance efficiency. CUO is integrated in Smart cloud Monitoring (SuMo), an open-source tool we develop for collecting monitoring data from Amazon Web Services (AWS) and analyzing them. A number of experiments are performed, using input data that correspond to realist AWS configuration scenarios, which exhibit the benefits of the CUO algorithm.