Using Multiple Seasonal Holt-Winters Exponential Smoothing to Predict Cloud Resource Provisioning

Abstract—Elasticity is one of the key features of cloud computing that attracts many SaaS providers to minimize their services’ cost. Cost is minimized by automatically provision and release computational resources depend on actual computational needs. However, delay of starting up new virtual resources can cause Service Level Agreement violation. Consequently, predicting cloud resources provisioning gains a lot of attention to scale computational resources in advance. However, most of current approaches do not consider multi-seasonality in cloud workloads. This paper proposes cloud resource provisioning prediction algorithm based on Holt-Winters exponential smoothing method. The proposed algorithm extends Holt- Winters exponential smoothing method to model cloud workload with multi-seasonal cycles. Prediction accuracy of the proposed algorithm has been improved by employing Artificial Bee Colony algorithm to optimize its parameters. Performance of the proposed algorithm has been evaluated and compared with double and triple exponential smoothing methods. Our results have shown that the proposed algorithm outperforms other methods. Keywords—auto-scaling; cloud computing; cloud resource scaling; holt-winters exponential smoothing; resource provisioning; virtualized resources I. INTRODUCTION Elasticity feature plays an important role in cloud computing by allowing SaaS providers to allocate and deallocate resources to their running services according to the demand. Elasticity allows SaaS providers to pay only for resources that are used by their cloud services [1]. However, the delay between requesting new resources and it being ready for use violates Service Level Agreement [2]. Therefore, forecasting future resource provisioning is needed to request resources in advance. Exponential Smoothing is a very popular smoothing method and has been used through years in many forecasting situations [3]. Many researchers have exploited Exponential smoothing methods to predict future resource provisioning for cloud computing applications [4][5]. However, most of them have used double exponential smoothing, which cannot model workloads if there are seasonalities. Most of cloud-computing applications’ workloads are influenced by seasonal factors (e.g., day, week, month, year) and have more than one seasonal pattern [6][7][8]. Workload has intraday seasonal pattern if there is a similarity of request when comparing requests of the corresponding hour from one day to the next day. Intraweek seasonal pattern exists if there is a similarity between requests in two corresponding days from two adjacent weeks [3]. Therefore, there is a strong demand to use predictive approach that is able to capture all seasonality patterns. This paper proposes resource usage prediction algorithm, which extends Holt-Winters exponential smoothing (HW) method to model multiple seasonal cycles. However, modeling multiple seasonal cycles requires large number of observation values. For example, predicting resource usage with intraday, intra-month, and intra-year seasonality patterns requires at least two years observation values. Moreover, finding optimal parameter values (smoothing constant, trend- smoothing constant and seasonal-smoothing constants) for multiple seasonality model is not an easy task. Therefore, the proposed algorithm detects seasonality patterns from available historical data by applying seasonality test, and extends HW accordingly to model detected seasonality patterns. While historical data size grows up and more seasonality patterns are detected, HW is gradually extended to be able to model detected seasonality patterns. Furthermore, prediction accuracy of the proposed algorithm has been enhanced by using artificial bee colony algorithm to find near optimal values for its parameters. Thus, unlike most of current resource prediction approaches, the proposed algorithm does not require any minimum number of observations values before applying it. However, good prediction accuracy will not be achieved until several steps have been made. The proposed algorithm has been evaluated using CloudSim simulator with real Web server log called Saskatchewan Log [6]. Performance of the proposed algorithm has been compared with double and triple exponential smoothing methods. Experimental results have shown that the proposed algorithm out

…(Full text truncated)…