Resource allocation within a mobile broadband (MBB) environment remains a challenging topic as it faces highly dynamic user demand and volatile network conditions while it must ensure an acceptable Quality of Experience (QoE) in the provided services without over-provisioning network resources. To solve this problem, we propose the implementation and experimental validation of a novel approach to intelligent resource allocation within MBB environments. Our approach enables the dynamic adaptation of resources to users’ demands effectively preventing possible degradation of the QoE of the service before it manifests to the client-side and becomes noticeable by the end users.
In our experiment, we validated our approach using a video streaming as the use case, which is one of the most popular and resource-demanding MBB services. Our methodology composes of 4 stages: (1) collection of metrics from the MONROE testbed and calculation of KPIs, (2) correlation of the KPIs to a measurable form of QoE, (3) prediction of future KPIs based on existing KPIs and the correlation model and finally (4) the generation of resource allocation actions both in the node (client-side) and in the server to improve the QoE. Our methodology includes state-of-the-art techniques for large-scale data correlation and employs machine-learning algorithms to forecast the degradation of QoE and the generation of suitable provisioning actions to prevent this degradation.
This report will present experimental results that demonstrate that the clustering and regression models implemented in the Qiqbus platform as well as the use of forecasting methods are effective in increasing the performance of a video session by enforcing appropriate provisioning actions which increase the Quality of Experience for the end user. Specifically, our Proof of Concept implementation answers the question on which provisioning action (server transmitting the video or interface being used by the container or both) should be used by MONROE containers within a given experiment environment. With provisioning actions provided as output from our clustering models, the experiments demonstrate the improvement of the Quality of Experience in various experiments both in testing mode and in deployed mode where a larger number of MONROE containers was employed so that we also validate the scalability of our Qiqbus decision engine when it handles multiple streams originating from many different containers.
From our business point of view, this project has provided us with important technical knowledge in defining a Qiqbus (www.modio.io/qiqbus_trailer) processing pipeline with appropriate machine learning methods (clustering and regression methods). We also managed to successfully integrate the Adaptive Data Series Index (ADS index) developed by partner Paris Descartes University, which performs in a scalable manner once the problem size becomes significantly large, which will be the case of the networked environment of a telecom operator with vast sets of devices using the substrate network. The results obtained in our experiment provide support for validating that our concept can be used to solve a number of networking challenges in dynamic mobile broadband environments that demand new solutions for dynamic network adaptation in order to increase the network performance and subsequently to improve the satisfaction of the end users of mobile broadband services. The knowledge that we gained in this project is helping us in two Future Internet Open Call projects which involves the implementation of a relevant processing pipeline for experimenting with a set of machine learning methods for solving the problem of autoscaling of highly dynamic network and cloud resources within mixed wired and wireless networked environments. We thus intend to continue use of the MONROE platform during the development of our Minimum Viable Product which is scheduled for the 1st quarter of 2019 with an estimated effort of 20 man months and which we will showcase to relevant industrial players and VCs with the goal to licence our technology, following the strong success record in recent major machine learning technology startup acquisitions, including DeepMind (Google), Magic Pony (Twitter) and Perceptio (Apple).