UNIVERSITY OF ZAGREB FACULTY OF ELECTRICAL ENGINEERING AND COMPUTING MASTER THESIS n. 721 PREDICTIVE MODEL OF FAILURES IN CLOUD COMPUTING Marija Mijić Zagreb, June 2014. To Mom, Dad and Ana Contents Introduction ........................................................................................................................... 1 1. Cloud Computing .......................................................................................................... 2 1.1. Essential Characteristics of Cloud Computing ...................................................... 4 1.2. Service Models ...................................................................................................... 6 1.3. Deployment Models .............................................................................................. 8 1.4. Current State of Cloud Computing ........................................................................ 9 1.5. Statistics ............................................................................................................... 11 2. Evaluation of the Cloud Computing Providers ........................................................... 12 2.1. Defining Cloud Performance ............................................................................... 12 2.2. Defining Availability and Failures in Cloud Computing .................................... 14 3. Real-time Tracking of the Availability of Cloud Computing Providers ..................... 16 3.1. Decentralized Cloud Solution ViFiB ................................................................... 16 3.2. Simple Language for Accounting and Provisioning operating system ................ 18 3.3. Re6st .................................................................................................................... 20 3.4. Availability Evaluation Model ............................................................................ 20 4. Experimental Evaluation of the Cloud Computing Providers and Failure Prediction 23 4.1. Evaluation with PyStone and SQLBench ............................................................ 23 4.2. Failure Prediction Model ..................................................................................... 25 4.2.1. Central Processing Unit Average Load vs. Usage ....................................... 27 4.2.2. Amazon ........................................................................................................ 28 4.2.3. Joyent ........................................................................................................... 30 4.2.4. Azure ........................................................................................................... 33 4.2.5. Rackspace .................................................................................................... 35 4.2.6. Comparison of Test Results and System Decision Making ........................ 37 Conclusion ........................................................................................................................... 38 References ........................................................................................................................... 39 Sažetak ................................................................................................................................. 41 Summary .............................................................................................................................. 42 Abbreviations ...................................................................................................................... 43 Appendix A: Detaljni opis stress-testiranja ......................................................................... 44 Appendix B: Detailed Description of the Stress Testing ..................................................... 45 Introduction Cloud computing industry has been growing since its creation. For many businesses this meant being able to fully focus on their core activities. It also meant that they would no longer have to buy expensive machines before they even needed them. Cloud computing is as much used by small companies as it is used by multinational corporations. And the transformation from a small business to a bigger one is supported by the cloud computing though its scalability. However, the current state of cloud computing, given the numerous outages, data loss and security breaches, does not justify its status of the future of information technology. One of the main problems of cloud computing lies in the fact that the cloud is centralized. Modern data centers are as big as factories and often difficult to supervise. This leaves cloud computing prone to attacks and unfit for fast recovery. In other words, the cloud is not resilient. Still, the statistics of the cloud failures are not widely known across the industry and the evaluation of the cloud still fails to generate a ranking list of cloud computing providers that would benefit both customers (when choosing the providers) and the providers (to know in which direction they should try to improve). In the Thesis, a new model of cloud computing will be used to improve its availability tracking and resilience. The first chapter explains some basic terms and concepts of the cloud computing. The second chapter contains the research done on the evaluation of current state of the cloud. It also provides an insight into the downtime statistics of the cloud computing providers as well as their ranking list. The third chapter explains how the evaluation of the cloud computing providers has been done so far. The fifth chapter proposes a new model of evaluation and ranking of the cloud computing providers. The last chapter shows the results of the testing of virtual machines of some of the biggest cloud computing providers today and explains how the results obtained from the tests can be used to rank and evaluate providers. 1 1. Cloud Computing Cloud computing is a model for enabling ubiquitous, convenient, on-demand network access to a shared pool of configurable computing resources (e.g., networks, servers, storage, applications, and services) that can be rapidly provisioned and released with minimal management effort or service provider interaction. This cloud model is composed of five essential characteristics, three service models, and four deployment models [1]. Many argue that the cloud had existed long before the actual term ‘cloud’ was coined. The cloud refers to both datacenter hardware and software and although the development of extremely large scale datacenters is considered to be the main reason of such wide use of the cloud today, there are other factors that played a big part in the global acceptance of cloud computing. Such factors, such as additional technology trends and new business models, made new application opportunities and usage models possible. Therefore, even though similar concepts existed, the usage was not quite the same as it is that of what is today considered to be cloud. Moreover, new application opportunities render cloud use more convenient than it has ever been. New Technology Trends and Business Models: 1. Credit cards accepted online Most of the cloud computing providers today offer scalable resources, “pay-as- you-go” model, which means that the cloud is used when needed and paid for when used. Such services could not be provided were it necessary to create a contract every time a client wanted to use more resources. In addition to that, small businesses that previously could not afford long-term commitments with payment processing service such as VeriSign or Authorize.net can now use low-cost card payment processing services such as PayPal. 2. Hardware-level virtualization The second innovation that helped the creation of cloud computing is the hardware-level virtualization that allows customers to choose their own stack without disrupting each other. That meant that the customers could now share the same hardware and, therefore further lower the costs. 2 New Application Opportunities 1. Mobile interactive applications Mobile phone sales grew 46.5 percent in second quarter of 2013 and exceeded feature phone sales for first time [2]. Mobile applications are now making up for as much as 86% of overall time spent on mobile device [3]. With mobile phone battery use being one of the biggest challenges and wide-area networking relatively higher than any other IT costs [4], mobile interactive applications, that require high availability and at the same time generate a large amount of data, will be able to use cloud to store data at a convenient distance. In addition to that, by removing some of the computation that is nowadays done by the phone itself and delegating them to the cloud, battery may last significantly longer. However, the traditional cloud model needs to be modified in order to adapt to these new services. Cloudlets are one of the suggested models for this type of use of the cloud. A cloudlet is a new architectural element that arises from the convergence of mobile computing and cloud computing. It represents the middle tier of a 3-tier hierarchy: mobile device --- cloudlet --- cloud. A cloudlet can be viewed as a "data center in a box" whose goal is to "bring the cloud closer". A Cloudlet is viewed in the very recent literature as a WiFi access-point used as an offloading facility for the mobile users. It opens the way to almost instantaneous mobile offloading. This property is of the highest importance for the mobile terminals manufacturers. 2. Parallel batch processing Cloud computing made fast data processing available even to small businesses. Contrary to the model where a company had to purchase costly machines that, in many cases, were going to be used occasionally, nowadays it is possible to leverage existing machines of other companies (such as Google’s App Engine or Amazon Elastic Compute Cloud) and pay only for the time machines were used. 3. The rise of analytics With parallel batch processing becoming so wide-spread, there is a rise in the analysis of the data and decision support systems that require the use of extensive computer resources. A use of such systems is found in trying to 3 understand customer behaviors supply chains, buying habits, ranking, and so on. 4. Extension of compute-intensive desktop applications Some of the most popular computation tools (such as Matlab and Mathematica) have integrated the cloud into their environment. It serves as an extension used to perform expensive evaluations. Other applications could possibly benefit from the cloud following the same path. The latest versions of the mathematics software packages Matlab and Mathematica are capable of using Cloud Computing to perform expensive evaluations [5]. 5. “Earthbound” applications There are, however, still some obstacles that render some applications inapt for user of the cloud. The cost of wide-area data transfer is still relatively hard, so, for applications that deal with a lot of data, it could be more expensive to move the data then to use their own equipment to do the computation. In addition, some applications require microsecond precision (for example, stock trading) and, considering the latency of wide-area data transfer, cloud cannot be used in such cases. 1.1. Essential Characteristics of Cloud Computing On-demand self-service. The contracts between a customer and a provider are automated with no need for human interaction. Similarly, the service can be stopped at any time, in case the customer decided they will not need the cloud anymore. Broad network access. The cloud is accessible over the network and can be accessed through think or thick client platforms (e.g., mobile phones, tablets, laptops, and workstations). Resource pooling. The providers’ resources are shared by multiple users using a concept called multitenancy. Multitenancy is a reference to the mode of operation of software where multiple independent instances of one or multiple applications operate in a shared environment. The 4