Regression analysis of predictions and forecasts of cloud data center KPIs using the boosted decision tree algorithm

Cloud data centers seek to optimize their provision of pooled CPU, bandwidth and storage resources. While over-provision is wasteful, under-provision may lead to violating Service Level Agreements (SLAs) with their consumers; yet the relationship between low-level Key Performance Indicators (KPIs) and SLA violations is not well understood. State-of-the art monitoring systems typically react to service failures after the fact, partly due to unexpected nonlinearities in the aggregated performance data. We seek to provide better modelling of KPIs using predictive algorithms that could be used for the proactive monitoring and adaptation of cloud services. In this paper, we investigate the Boosted Decision Tree (BDT) regression algorithm. We tested the BDT algorithm in a real monitoring framework deployed on a novel Azure cloud test-bed distributed over multiple geolocations, using thousands of robot-user requests to produce huge volumes of KPI data. The BDT algorithm achieved an R-Squared score of 0.9991 at the 0.2 learning rate. This closely predicted the KPI data and outperformed other approaches, such as Ordinary Least Squares and Stochastic Gradient Descent; and is a promising candidate for making short- and long-term predictions for cloud resource allocation.