Auto-Scaling Provision Basing on Workload Prediction in the Virtualized Data Center

Introduction

Recently, the two main features of the cloud computing are both the elasticity (Herbst, Kounev, & Reussner, 2013) and virtualization (Yang et al., 2009) in the cloud data centers. That is to say, it is important to meet the fluctuating demand (Weingärtner, Bräscher, & Westphall, 2015) of the users. However, static lower or upper threshold would lead the poorer utilization and the more energy consumption. If the resources are under-provisioning, it would not meet the demand of the users, and cause the punishment by the SLA. If the resources are over-provisioning, it is another emergent problem to solve for the energy consumption (Chihoub et al., 2015) and heat loss.

To make the further fine-grained scaling in the resource provisioning, the auto-scaling provisioning (Roy et al., 2011) used to be combined with the predict technique. The purpose of it is to avoid excessive or inadequate resources to be allocated. The under-provisioning resources would not meet the demand of the users, which causes SLA violations (Musa et al., 2016). On the contrary, the over-provisioning resources would cause the resources waste during the deploying process in the cloud data centers. Thus, the two challenges in the allocating resources of the cloud computing is to solve the following problem: (1) deciding the efficient elastic provisioning (Dustdar et al., 2011) in the data centers, (2) improving the prediction accuracy and reducing the time complexity, (3) minimizing the renting cost and service level objective (SLO) violations (Manvi & Shyam, 2014).

Therefore, in this paper, we propose a novel combined prediction technique as a kind of efficient resource provisioning strategy in the cloud data centers, as it is shown in Figure 1. The combined prediction algorithm is mainly composed of two parts. The basic algorithm is the ES model with PSO algorithm due to its flexible calculation basing on a few samples. Then, the reactive prediction model is PF model just for reducing the under-provisioning statements. The aim of the combined prediction model is to minimize the renting cost and the SLO violations (Hwang et al., 2013). We analyze and determine the features of the demand which is composed of repeated patterns or unrepeated ones. The purpose of the proposed model is to improve the accuracy and minimize the overheads. According to the analysis, we present the two-level forecasting technique:

• •

  In the first level, we choose the ES model to predict the varying demand with the workload. Single exponential smoothing curve is fit to forecast closer to actual observations by setting different weights. Then, the weight is based on the calculation of the prediction accuracy. If the time series have the rising or falling trend, the prediction deviation would be less prediction accuracy, then it would get a better result when combining the exponential smoothing method with the reasonable heuristic algorithm for the further forecast. Actually, PSO algorithm could achieve the suitable result more quickly. Considering the fluctuation of the workload with the time, we propose the ES prediction combining with the particle swarm optimization (PSO) to make the higher prediction accuracy. For the further step, the PSO would quickly search and determine the suitable value by the reasonable fitness function in order to meet the varying demand of the users;

• •

  In the second level, considering the random variation during the workload, we address the simple Polynomial Fitting (PF) model to adjust the irregular demand when the workload changes with the time series. The purpose of the PF model is making efforts to reduce the under-provisioning statements. As the reactive prediction basing the ES algorithm, it would avoid the under-provisioning statement efficiently. For the future work, as a reactive strategy, it would avoid the SLA violations due to reactive demand of the users;

• •

  The core problem of the combination prediction model is to decide the reasonable prediction model. Actually, the fluctuating workload is periodical and random. Some workload curves are periodical obviously. At the same time, we also consider the random of the workload. Then, the reactive prediction technique is the supplement to avoid the under-provisioning statement. The combined prediction model considers not only the similarity of the workload, but also the variable curve among the workload.