دانلود رایگان ترجمه مقاله پیش بینی طوفان در یک ابر (نشریه IEEE 2013)

پیش بینی طوفان در یک ابر

Storm Prediction in a Cloud

Abstract

Predicting future behavior reliably and efficiently is key for systems that manage virtual services; such systems must be able to balance loads within a cloud environment to ensure that service level agreements are met at a reasonable expense. In principle accurate predictions can be achieved by mining a variety of data sources, which describe the historic behavior of the services, the requirements of the programs running on them, and the evolving demands placed on the cloud by end users. Of particular importance is accurate prediction of maximal loads likely to be observed in the short term. However, standard approaches to modeling system behavior, by analyzing the totality of the observed data, tend to predict average rather than exceptional system behavior and ignore important patterns of change over time. In this paper, we study the ability of a simple multivariate linear regression for forecasting of peak CPU utilization (storms) in an industrial cloud environment. We also propose several modifications to the standard linear regression to adjust it for storm prediction.

1 Introduction

Infrastructure as a Service (IaaS) is becoming a norm in large scale IT systems and virtualization in these environments is common. One of the main difficulties of such virtualization is the placing of virtual machines (VMs) and balancing the load. If the demands placed on the infrastructure exceed its capabilities, thrashing will occur, response times will rise, and customer satisfaction will plummet. Therefore it is essential to ensure that the placing and balancing is done properly [1-4].

Proper balancing and capacity planning in such cloud environments requires forecasting of future workload and resource consumptions. Without good forecasts, cloud managers are forced to over-configure their pools of resources to achieve required availability, in order to honor service level agreements (SLAs). This is expensive, and can still fail to consistently satisfy SLAs. Absent good forecasts, cloud managers tend to operate in a reactive mode and can become ineffective and even disruptive.

Several workload forecast techniques based on time series analysis have been introduced over the years [5] that can be applied in the cloud settings as well. The bottom-line of such literature is that there is no “silver bullet” technique for forecasting. Depending on the nature of the data and characteristics of the services and the workload, different statistical techniques and machine learning algorithms may perform better than the others. In some cases even the simplest techniques such as linear regression may perform better than the more complex competitors [6].

To understand the practicality of such prediction techniques on industrial size problems, we set up a series of case studies where we apply different forecasting techniques on data coming from our industrial collaborator, CA Technologies [7]. CA Technologies is a cloud provider for several large scale organizations. They provide IaaS to their clients and monitor their usage. Their cloud manager system basically is responsible for balancing the workload by placing the virtual machines on the physical infrastructure.

In this paper, we report our experience on applying a basic multivariate linear regression (MVLR) technique to predict the CPU utilization of virtual machines, in the context of one of the CA clients. However, unlike many existing prediction techniques, where they minimize the average prediction errors or maximize average likelihoods, we are more interested in predicting extreme cases rather than averages. The motivation comes from the type of workload we are facing in our case study, which is not very uncommon for other cloud-based applications, as well. In our case, the average utilization across all VMs was at most 20%, but the maximum utilization was almost invariably very close to 100%. Applying MVLR in such data (most of the time very low utilization but occasionally reaching to peaks), we realized that though the average predictions are very accurate but the forecast for large values (storms) are drastically poor.

To cope with this problem, we introduce several modifications to the basic MVLR to adjust it for predicting peak values. The results show that subtracting seasonalities extracted by Fourier transform and then using a weighted MVLR provides our best observed results for storm prediction. In the following sections, we describe the details of each modified MVLR and report its results.