Stochastic Scheduling Informed by Probabilistic Forecasts of Computing Resource Requirements
Open AccessCloud computing has emerged as a dominant paradigm, where a colossal scale serves as the claim to fame. Scheduling exists as a key component of cloud computing, and small improvements in efficiency can save millions of dollars for large companies. The use of machine learning algorithms to inform traditional scheduling algorithms remains unexplored, and our research proposes a method to increase data center utilization 10% by providing forecasts to the scheduler.Several researchers have attempted to forecast the demand for cloud services. However, no previous studies have incorporated the probabilistic component of forecasts into the scheduling procedure. In efforts to develop a proof-of-concept application to schedule computing tasks based on probabilistic forecasts, we reviewed relevant literature related to forecasting computing requirements for tasks. Then we analyzed our data to find the most relevant features and selected various machine learning algorithms to determine which methods yield the best performance. Prior CPU and memory utilization measures served as key features to predict future utilization, but some covariates also played a role in forecasting the computing requirements. After training the models, we back-tested the models against a holdout set to evaluate performance. We found reducing the number of available machines by 10% avoided out of memory errors and safely increased data center utilization by 11%. While one must evaluate the approach in a live setting, the findings provide a foundation for continued research.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.