Workload Prediction on Google Cluster Trace

Abstract

Workload prediction in cloud systems is an important task to ensure maximum resource utilization. So, a cloud system requires efficient resource allocation to minimize the resource cost while maximizing the profit. One optimal strategy for efficient resource utilization is to timely allocate resources according to the need of applications. The important precondition of this strategy is obtaining future workload information in advance. The main focus of this analysis is to design and compare different forecasting models to predict future workload. This paper develops model through Adaptive Neuro Fuzzy Inference System (ANFIS), Non-linear Autoregressive Network with Exogenous inputs (NARX), Autoregressive Integrated Moving Average (ARIMA), and Support Vector Regression (SVR). Public trace data (workload trace version II) which is made available by Google were used to verify the accuracy, stability and adaptability of different models. Finally, this paper compares these prediction models to find out the model which ensures better prediction. Performance of forecasting techniques is measured by some popular statistical metric, i.e., Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Sum of Squared Error (SSE), Normalized Mean Squared Error (NMSE). The experimental result indicates that NARX model outperforms other models, e.g., ANFIS, ARIMA, and SVR.