Forecasting Cloud Resource Utilization Using Time Series Methods

University essay from KTH/Skolan för elektroteknik och datavetenskap (EECS)

Author: Prashant Kumar; [2018]

Keywords: ;

Abstract: With the contemporary technological advancements, the adoption of cloud as service has been evolving exponentially while providing a seemingly incessant measure of resources such as storage, network, CPU and many more. In today’s data centres that accommodate thousands of servers, ensuring the availability of continuous services is a significant hurdle. So as in to meet the compelling demands of resources, the proactive forecasting of resource usage demands is of immense importance. Cloud service providers can profit by having an unbiased forecast of the future usage requirement on a cloud resource proactively by employing historical data patterns. Therefore, forecasting resource usage is of great importance for dynamic scaling of cloud resources to have gained such as cost saving and optimal energy consumption while vouching for the proper quality of service. Amongst the extensive resources in a cloud setup, we focus on CPU utilisation metric. In this work, we review the performance of several forecasting methods discussed in existing literature, experiment with ensemble models, recognise the fundamental evaluation metrics, reform data to a machine learning common problem and ultimately compare the performance of the model on the given dataset. To assess the accuracy of the forecasting model we validate it against the unseen test data using walk-forward validation technique. As a conclusion, we found the Feed Forward Neural Network to be the best accomplishing model when evaluated with real traces of CPU utilisation of a cloud setup where it showed an improvement of approximately 18.13% when relative measure of forecasting error is considered. We also find that an amalgam of individual forecasting models such as ARIMA and ETS performs better with an improvement of approximately 2.6% than the individual time series method such as ARIMA in our case. In the end, we also discuss the possible approaches which could improve the performance of this work and the possible future work to encourage fur-ther research in the area of time series forecasting.

  AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)