**6. Discussion**

Among the different aspects to be taken into consideration in a smart classroom [14], the proposed architecture focuses on orchestrating the complex technical ecosystem and enabling its "smart" features. The architecture has been designed bearing the following main requirements in mind: within-scenario and between-scenario flexibility, seamless privacy and authentication configurations, and easy communication with external data sources.

The experimental results regarding the performance and scalability of our architecture show how heterogeneous classroom devices can be managed in an automatic and efficient way to host different amounts and types of learning tools and applications. Concretely, we demonstrated the scalability of our architecture when an increasing number of Dockers, with diverse computational requirements, is deployed over three widely used hardware configurations such as laptops, personal computers and servers. However, no direct comparison of the obtained results with those reported in the literature was possible since they highly depend on the hardware and software configuration. Furthermore, most MMLA studies evaluate their results based on educational outcomes but not on technical performance.

The automatic and flexible managemen<sup>t</sup> of the proposed architecture has been motivated through the case study presented in this paper, which illustrates the limitations of current solutions and how our proposal offers a seamless switch between three different learning scenarios happening in the same smart classroom. While existing architectures for smart classrooms often involve ad-hoc digital devices and tools that can be used in specific ways [15,22], in our proposal, the different modules of the ecosystem can be orchestrated for multiple purposes in scalable and interoperable ways. Moreover, the human intervention required to adapt and reconfigure the transition between heterogeneous learning lessons is significantly reduced and can be automatized.

The presented architecture could be of grea<sup>t</sup> value also for the remote lab community. While virtualization techniques had been already explored [27–29], this architecture could increase the flexibility of remote labs, by supporting the configuration and deployment of remote experiments [32]. Moreover, it supports the collection of multimodal data (coming both from hardware and software) necessary to support the smart adaptation to the learning process.

Regarding the instant and adaptive support expected from smart classrooms [13], our proposal could become the base upon which other architectures could build, uncoupling the multimodal challenges of the DVC [4,63]. More concretely, our contribution helps to address the lower level technical requirements of the DVC, and more conceptual architectures (e.g., [21,22,63]) could build on top of it. Thus, our proposal contributes to diminishing the need for ad-hoc MMLA solutions often due to the technical constrains to the ecosystem [8]. As a consequence, relying on a lower level architecture will open the door to multiple analysis and adaptability schemes in smart classrooms, addressing the reusability and interoperability problems among MMLA solutions [9,10].

The integration of SDN/NFV in our architecture allows instructors to reduce their workload avoiding the manual configuration of classroom devices according to the topic and purpose of each subject. It also reduces the complexity of the smart classrooms managemen<sup>t</sup> as well as optimizes the usage of classrooms devices. In a nutshell, smart classrooms equipped with our architecture will be able to reconfigure and optimize the learning applications of their devices ant their communications according to the current subject topic and number of students. It will be done in real-time and on-demand. In contrast, as it has been demonstrated in Section 2, existing solutions using virtualization techniques [25,26] are not able to reconfigure the whole remote lab in a flexible way. They just consider predefined VMs implementing particular learning applications that are instantiated and dismantled. It means that they miss critical aspects such as the flexible managemen<sup>t</sup> of communications, essential to guarantee QoS issues when the number of students increases, and the optimization of hardware resources of learning devices such as CPU, memory and storage.

It is important to note that one of the main limitations of the proposed architecture is the complexity of its deployment. The usage of resource-constrained devices such as digital boards or cameras makes very complex their managemen<sup>t</sup> through current virtualization techniques. Fortunately, this issue is limited when other devices such as tablets and personal computers are considered in smart classrooms. Additionally, the architecture is still to be tested in a real scenario, which is part of the future work. Moreover, we argue that the architecture represents an improvement with respect to other studies. However, we cannot present a direct comparison in terms of efficiency because most MMLA studies do not report on the performance of the architectures from the technical point of view. Finally, we still have not tackled the challenge of how instructors will be able to interact with this architecture through a user-friendly authoring tool.

#### **7. Conclusions and Future Directions**

Smart classrooms require a dynamic and flexible orchestration of their complex ecosystem, currently performed manually by instructors that use ad-hoc learning applications. With that goal in mind, this paper the following three key research problems: (1) the limitations of current learning solutions in terms of flexible and scalable managemen<sup>t</sup> of devices belonging to simulated and realistic learning scenarios; (2) the suitability of technologies and their integration in an architecture able to provide the level of flexibility and dynamicity required by current learning environments; and (3) the scalability and performance of the architectures. With challenges in mind, this paper proposes an MEC-enabled architecture that considers SDN/NFV to reconfigure the software and hardware resources of classroom devices in real-time and on-demand. A case study inspired by authentic learning analytics applications extracted from the literature has been proposed to highlight the limitation of the existing solution and demonstrate the added value of our architecture. The experimental results demonstrate acceptable computational performance and efficiency when typical classroom devices such as servers, personal computers or laptops implementing practical learning tools are deployed and reconfigured. Specifically, we investigated experiments with different MEC Apps such as face detector, ASR and physics simulation, each one with different computational requirements. The results point out the potential of our architecture to manage heterogeneous classroom devices in an automatic and efficient way.

As future work, we plan to implement and deploy the proposed architecture in a realistic smart classroom scenario to demonstrate its usefulness with real students. In this sense, we will integrate our architecture in existing platforms able to deploy, dismantle and control the life-cycle of VMs and containers such as OpenStack, as well as control the network infrastructure and the communications of the smart classroom by using OpenDaylight as SDN Controller.

**Author Contributions:** Conceptualization, A.H.C., J.A.R.-V., F.J.G.C. and M.J.R.-T.; Funding acquisition, G.M.P.; Methodology, A.H.C., J.A.R.-V. and F.J.G.C.; Resources, G.M.P.; Software, F.J.G.C.; Supervision, J.A.R.-V.; Visualization, A.H.C. and F.J.G.C.; Writing—original draft, A.H.C., J.A.R.-V., F.J.G.C., M.J.R.-T. and S.K.S.; Writing—review and editing, A.H.C., J.A.R.-V., F.J.G.C., M.J.R.-T., S.K.S. and G.M.P. All authors have read and agreed to the published version of the manuscript.

**Acknowledgments:** This work has been partially supported by the Government of Ireland post-doc fellowship (grant code GOIPD/2018/466 of the Irish Research Council), the Spanish Ministry of Economy and Competitiveness through the Juan de la Cierva Formación program (FJCI-2017-34926), and the European Union via the European Regional Development Fund and in the context of CEITER (Grant agreements No.669074).

**Conflicts of Interest:** The authors declare no conflict of interest.
