ACCORDION - D2.2 State of the art report (I)

Korontanis I., Carlini E., Kavalionak H., Dazzi P., Psomakelis V., Blasi L., Nadir Z., Violos J., Toro A., Russo M., Fabregas E. M., Vasios G., Tserpes K., Zadtootaghaj S., Lipa B., Pateraki M., Di Girolamo M.

Edge  Resource pooling  Orchestration  Machine learning  Models 

The present document is the result of the collaborative effort of all ACCORDION partners participating to Task 2.2. The document offers a review of the state of the art for a series of topics strictly related to the work performed in the ACCORDION project. There is actually a strict correlation between the topics analyzed in this document and the Tasks that are part of the three research Work Packages of ACCORDION (WP3, WP4, and WP5). The main part of this document is section 2, in which all the state of the art analysis results have been reported. Section 2 has a subsection for each of the topics researched in the project, which includes: a description of the objectives, a list of outcomes expected from the research work, and an analysis of the state of the art. Section 2.1 (Resource monitoring & characterization) reports on monitoring, characterization and classification of Edge resources, identifying Prometheus, TOSCA and the automatic creation of taxonomies, respectively, as the best solutions for each of the three fields. Section 2.2 (Resource indexing & discovery) focuses on discussing solutions and data structures for organizing data in Resource Discovery Services. Section 2.3 (Edge storage, availability, reliability and performance) presents the advantages and disadvantages of both block and object storage, and then discusses some solutions, identifying OpenStack and MinIO as the most promising ones, even if not completely suitable. Some open research issues are also summarized. Section 2.4 (Pooling Edge resources), after listing some orchestration challenges typical of Edge computing and the techniques to cope with them, reports on several solutions to be considered as possible baselines for the ACCORDION Minicloud. Section 2.5 (??-based network orchestration) first lists the main machine learning techniques, then explores both Federated Learning techniques and further evolutions such as Meta-Learning Framework and Multi-Agent Reinforcement Learning. Section 2.6 (Resilience policies & mechanisms over heterogeneous edge resources) starts by discriminating between reactive and proactive protection strategies and describing some of them. Then other Fault Tolerance approaches are explored both reported in the literature and adopted in common distributed computing frameworks (Openstack, Cloudstack, Kubernetes, Openshift, and Mesos). Finally techniques for movement behaviour and resource utilization prediction are analysed, with a particular focus on the the LSTM model for Neural Networks. The conclusion is that the most promising solution to efficiently adapt the deep learning topologies for the fault tolerance needs is the hyper parameter optimization approach. Section 2.7 (Techniques for secure Edge application development & deployment) offers an analysis of the most common types of security attacks (Distributed Denial-of-Service, Malware Injection, and Authentication-based attacks) and their related countermeasures, along with some threat modelling methods, while DevSecOps methods and tools are also described. Section 2.8 (Privacy preserving mechanisms) starts by analysing Machine Learning techniques with a focus on privacy preserving ones, and then lists a number of works analysing how cookie synchronization techniques adopted for web advertising can expose users to privacy leaks. Section 2.9 (Application model for automatic deployment / migration of components) looks for application description models suitable for ACCORDION, i.e. with a machine-processable syntax, able to represent resource capacity requirements, containerization, and recovery policies. Three available solutions, TOSCA, Juju charms and CAMP, are compared along with the projects that are using them. Furthermore, tools supporting the three above solutions are described, and works researching the interoperability among the solutions are also analysed. Section 2.10 (Modelling and assessing QoE for NextGen applications) reports on different types of objective models that can be used to estimate the Quality of Experience (QoE) perceived by users of multimedia applications, and about the latest ITU-T Recommendations on QoE models and methodologies that can be applied to Next Generation Applications. For the ACCORDION project it has been decided to follow the standardized approach to build models for QoE assessment of ACCORDION applications. Section 2.11 (DevOps tools to automate Edge applications' deployment) sets the context and reports the starting points for the evaluation of Continuous Integration and Continuous Deployment tools. The identified state-of-the art solutions are Jenkins for the CI/CD pipeline and Kubernetes as the runtime deployment environment. Section 2.12 (Collaborative VR), starting from the general requirements for Virtual Reality applications, reports considerations about the still limited power of the available HMDs and discusses the trade-offs conditioning the possibility to offload computation from the end devices to the Edge. Finally Section 2.13 (Resource federation models) describes the main features of the federation model proposed by the H2020 5GeX project and lists the additional constraints and issues raised by an Edge providers' federation, which have to be further investigated. Not all project's research Tasks have a related section in this document about their main topics, yet. Monitoring the State of the Art is an ongoing activity in ACCORDION, and the next version of this document foreseen in M22 will improve the State of the Art analysis by adding further details, covering more topics and reporting on possible new approaches that appeared in the meantime.

Source: ISTI Project report, ACCORDION, D2.2, 2020