ICON supports the localized, personalized and timely provision of media resources to users through the adoption of the FLAME’s platform. End users can search and consume content based on their preferences learned by the provided ICON application. At the same, ICON supports the concept of content prosumers, i.e., end users can become content producers, apart from being consumers, having the opportunity to upload media files into the available infrastructure, thus, setting up the basis for novel business models.
ICON, which has been designed, developed, integrated and tested with the FLAME platform over the course of 6 months (January – June 2020) uses machine learning algorithms to offload video content to the EDGE before the user request takes place and utilizes content caching strategies to adapt to the current network conditions. Special pre-trained feed-forward and deep neural network models are deployed that utilize video characteristics such as popularity and size and also the location of the uploaded video in order to choose whether the content should be offloaded to the EDGE. In order to decide the amount of time a video should stay stored at the EDGE the K-means machine learning algorithm is employed that dynamically adapts the caching strategy according to video sizes and the available storage resources at the Edge.
In order to integrate ICON with the FLAME platform and make use of its features, the necessary development and integration was performed. After the integration, testing in the Sandpit showed that the key functionality was operative and only minor tweaks were necessary to carry the solution to the Barcelona testbed. At this point, a series of KPIs already used to determine the performance of ICON and to validate the solution: The service response time, service reliability, user bandwidth, download times, Edge serving ratio. Also, in preparation for the trial and the physical experiments in Barcelona, a web application was adopted running in browser in users’ devices, undertaking the responsibility of maintaining the video database while communicating with the available edge access points (i.e., FLAME’s clusters per access point).
During the trial, which was performed with several users on-street, it was observed that after a 15 minute period of users accessing video content, the machine learning algorithms managed to proactively offload the most of the content to the Edge. As a result, 88-90% of the users were served by the Edge server and thus, experiencing very short download/upload times, with at least 99.99% of service reliability. We consider such results very satisfactory and we are confident they highlight the efficiency of the ICON solution.