CPU workload forecasting of machines in data centers using LSTM recurrent neural networks and ARIMA models

Abstract

The advent of Data Science has led to data being evermore useful for an increasing number of organizations who want to extract knowledge from it for financial and research purposes. This has triggered data to be mined at an even faster pace causing the rise of Data Centers that host over thousands of machines together with thousands of jobs running in each of those machines. The growing complexities associated with managing such a huge infrastructure has caused the scheduling management systems to be inefficient at resource allocation across these machines. Hence, resource usage forecasting of machines in data centers is a growing area for research. This study focuses on the Time Series forecasting of CPU usage of machines in data centers using Long Short-Term Memory (LSTM) Network and evaluating it against the widely used and traditional autoregressive integrated moving average (ARIMA) models for forecasting. The final LSTM model had a forecasting error in the range of 17–23% compared to ARIMA model's 3742%. The results clearly show that LSTM models performed more consistently due to their ability to learn non-linear data much better than ARIMA models.