2 September 2024 in IEEE Network
Extended Reality (XR) refers to a class of contemporary services that are intertwined with a plethora of rather demanding Quality of Service (QoS) and functional requirements. Despite Kubernetes being the de facto standard in terms of deploying and managing contemporary containerized microservices, it lacks adequate support for cross-cluster networking, hindering service-to-service communication across diverse cloud domains. Although there are tools that may be leveraged alongside Kubernetes in order to establish multi-cluster deployments, each one of them comes with its own drawbacks and limitations. The purpose of this article is to explore the various potential technologies that may facilitate multi-cluster deployments and to propose how they may be leveraged to provide a cross-cluster connectivity solution that caters to the intricacies of XR services. The proposed solution is based on the use of two open-source frameworks, namely Cluster API for multi-cluster management, and Liqo for multi-cluster interconnectivity. The efficiency of this approach is evaluated in the context of two experiments. This work is the first attempt at proposing a solution for supporting multi-cluster deployments in a manner that is aligned with the requirements of XR services.
30 April 2026 in IEEE Transactions on Green Communications and Networking
The proliferation of computation-intensive services has raised growing concerns regarding the sustainability of future computing infrastructures. This paper investigates distributed edge computing systems powered by hybrid energy sources, where we jointly consider energy consumption from computation, cooling, and power source switching. To improve sustainability, we formulate an optimization problem that aims to minimize conventional energy consumption by jointly coordinating task distribution and power source switching. The problem involves complex resource constraints, heterogeneous dynamics, partial observability, and coupled task-power decisions, rendering conventional optimization methods ineffective. To address this challenge, we decompose the problem into task allocation, task selection, and power switching subproblems, following the system execution logic and constraints. Then, we propose an end-to-end solution based on heterogeneous cooperative multi-agent deep reinforcement learning, in which dedicated agents are designed to address individual subproblems while maintaining system coordination. Extensive simulations results demonstrate that the proposed approach achieves enhanced coordination and substantial reductions in conventional energy consumption, resulting in improved overall system sustainability.
14 June 2018 in IEEE Transactions on Services Computing
Due to their heavy reliance on server infrastructure, the current computation-intensive SaaS suffer from scalability issues compared to the existing data-intensive commercial SaaS. Offloading certain computations to the client browser can resolve these scalability issues but the current Browser APIs are complex to use and integrate in a single software. We propose in this paper four high level APIs that harness existing browser-based paradigms and proven software architectures to reduce the complexity of parallel computing and device-agnostic interactive rendering in the web browser. To allow experimental results, we have developed a proof-of-concept browser-based and interactive diffusion MRI where we have particularly deployed a parallel Diffusion Tensor Estimation. Our platform provides us easy APIs achieving up to 4 times speedup in parallel computation and real-time interactive rendering performances across different Browsers and Devices in comparison to other existing solutions in the Diffusion MRI Community.