Predictive model creation approach using layered subsystems quantified data collection from LTE L2 software system

• Main Author: Puerto Valencia, J. (Jose)
• Format: Dissertation
• Language: English
• Published: University of Oulu 2019
• Online Access: http://jultika.oulu.fi/Record/nbnfioulu-201907192705

Abstract. The road-map to a continuous and efficient complex software system’s improvement process has multiple stages and many interrelated on-going transformations, these being direct responses to its always evolving environment. The system’s scalability on this on-going transformations depends, to a great extent, on the prediction of resources consumption, and systematic emergent properties, thus implying, as the systems grow bigger in size and complexity, its predictability decreases in accuracy. A predictive model is used to address the inherent complexity growth and be able to increase the predictability of a complex system’s performance. The model creation processes are driven by the recollection of quantified data from different layers of the Long-term Evolution (LTE) Data-layer (L2) software system. The creation of such a model is possible due to the multiple system analysis tools Nokia has already implemented, allowing a multiple-layers data gathering flow. The process starts by first, stating the system layers differences, second, the use of a layered benchmark approach for the data collection at different levels, third, the design of a process flow organizing the data transformations from recollection, filtering, pre-processing and visualization, and forth, As a proof of concept, different Performance Measurements (PM) predictive models, trained by the collected pre-processed data, are compared. The thesis contains, in parallel to the model creation processes, the exploration, and comparison of various data visualization techniques that addresses the non-trivial graphical representation of the in-between subsystem’s data relations. Finally, the current results of the model process creation process are presented and discussed. The models were able to explain 54% and 67% of the variance in the two test configurations used in the instantiation of the model creation process proposed in this thesis.