Time Series Forecasting using Facebook Prophet for Cloud Resource Management

The heterogeneous nature of workloads running in cloud environments makes future resource usage prediction a complicated problem. Virtual machines can be described in five types of resource utilization patterns: steady, trending, seasonal, cyclic, and bursty behavior. Understanding these usage patterns and behaviors can enhance resource management on cloud data centers, especially VM scheduling, power management, and server health management systems. This paper applies the Facebook Prophet forecast framework on Microsoft Azure VM workload to predict future resource utilization required by the running tasks. We conclude that utilizing data preprocessing and transformation on real virtual machine traces, and incorporating an automatic model hyperparameter tuning process, can significantly increase forecasting accuracy with an average percentage change of over 85%. Furthermore, cloud providers can learn from their data center workloads and employ various forecasting models to gain substantial improvements in cost-efficient resource management.

Automated summary

Utilizing data preprocessing and transformation on real virtual machine traces, combined with automatic model hyperparameter tuning, significantly increases forecasting accuracy for resource utilization. Cloud providers can learn from their workloads and employ various forecasting models to achieve substantial improvements in cost-efficient resource management.