To achieve “provisioning elasticity”, the cloud needs to manage its available resources on demand. A-priori, static, VM provisioning introduces no runtime overhead but fails to handle unanticipated changes in resource demands. Dynamic provisioning addresses this problem but introduces runtime overhead. To avoid sub-optimal provisioning my PhD thesis adopts a hybrid approach that combines static and dynamic provisioning. The idea is to adapt an initial static placement of VMs in response to evolving load characteristics. My work is focused on broadening the applicability of clouds by looking at how the infrastructure can be more effectively used to support historically atypical applications (e.g. those that are interactive in nature with tighter QoS constraints). To accomplish this I have developed a family of related algorithms that collectively improve resource sharing on physical machines to permit load variation to be better addressed and to lessen the probability of VM interference due to resource contention. The family includes three core dynamic provisioning algorithms. The first algorithm provides for the short-term, controlled sharing of resources between co-hosted VMs, the second identifies pairs (and by extrapolation larger groups) of VMs that are predicted to be "compatible" in terms of the resources they need. This allows the cloud provider to do co-location to make the first algorithm more effective. The final, third, algorithm deals with under-utilized physical machines by re-packing the VMs on those machines while also considering their compatibility. This final algorithm both addresses the possibility of the second algorithm creating underutilized machines as a result of pairing and migration and also handles underutilization arising from “holes” left by the termination of short-duration VMs (another form of atypical VM application). I have also created a surprisingly simple static provisioning algorithm that considers compatibility to minimize VM interference that can be used before my dynamic algorithms. My evaluation is primarily simulation-based though I have also implemented the core algorithms on a small test-bed system to ensure correctness. The results obtained from my simulation experiments suggest that hybrid static and dynamic provisioning approaches are both feasible and should be effective supporting a broad range of applications in cloud environments. / February 2017

Cloud computing setups are a huge investment of resources and personnel to maintain. As<br />the workload on a system is a major contributing factor to both the performance of the<br />system and a representation of the needs of system users, a clear understanding of the<br />workload is critical to organizations that support supercomputing systems. In this paper,<br />we analyze traces from two production level supercomputers to infer the characteristics of<br />their workloads, and make observations as to the needs of supercomputer users based on<br />them. We particularly focus on the usage of graphical processing units by domain<br />scientists. Based on this analysis, we generate a synthetic workload that can be used for<br />testing future systems, and make observations as to e"cient resource provisioning. / Master of Science

The rate of data growth in many domains is straining our ability to manage and analyze it. Consequently, we see the emergence of computing systems that attempt to efficiently process data-intensive applications or I/O bound applications with large data. Cloud computing offers “infinite” resources on demand, and on a pay-as-you-go basis. As a result, it has gained interest for large-scale data processing. Given this supposedly infinite resource set, we need a provisioning process to determine appropriate resources for data processing or workload execution. We observe that the prevalent data processing architectures do not usually employ provisioning techniques available in a public cloud, and existing provisioning techniques have largely ignored data-intensive applications in public clouds. In this thesis, we take a step towards bridging the gap between existing data processing approaches and the provisioning techniques available in a public cloud, such that the monetary cost of executing data-intensive workloads is minimized. We formulate the problem of provisioning and include constructs to exploit a cloud’s elasticity to include any number of resources to host a multi-tenant database system prior to execution. The provisioning is modeled as a search problem, and we use standard search heuristics to solve it. We propose a novel framework for resource provisioning in a cloud environment. Our framework allows pluggable cost and performance models. We instantiate the framework by developing various search algorithms, cost and performance models to support the search for an effective resource configuration. We consider data-intensive workloads that consist of transactional, analytical or mixed workloads for evaluation, and access multiple database tenants. The workloads are based on standard TPC benchmarks. In addition, the user preferences on response time or throughput are expressed as constraints. Our propositions and their results are validated in a real public cloud, namely the Amazon cloud. The evaluation supports our claim that the framework is an effective tool for provisioning database workloads in a public cloud with minimal dollar cost. / Thesis (Ph.D, Computing) -- Queen's University, 2013-11-30 19:30:39.427

Big data has received a momentum from both academia and industry. The MapReduce model has emerged into a major computing model in support of big data analytics. Hadoop, which is an open source implementation of the MapReduce model, has been widely taken up by the community. Cloud service providers such as Amazon EC2 cloud have now supported Hadoop user applications. However, a key challenge is that the cloud service providers do not a have resource provisioning mechanism to satisfy user jobs with deadline requirements. Currently, it is solely the user responsibility to estimate the require amount of resources for their job running in a public cloud. This thesis presents a Hadoop performance model that accurately estimates the execution duration of a job and further provisions the required amount of resources for a job to be completed within a deadline. The proposed model employs Locally Weighted Linear Regression (LWLR) model to estimate execution time of a job and Lagrange Multiplier technique for resource provisioning to satisfy user job with a given deadline. The performance of the propose model is extensively evaluated in both in-house Hadoop cluster and Amazon EC2 Cloud. Experimental results show that the proposed model is highly accurate in job execution estimation and jobs are completed within the required deadlines following on the resource provisioning scheme of the proposed model. In addition, the Hadoop framework has over 190 configuration parameters and some of them have significant effects on the performance of a Hadoop job. Manually setting the optimum values for these parameters is a challenging task and also a time consuming process. This thesis presents optimization works that enhances the performance of Hadoop by automatically tuning its parameter values. It employs Gene Expression Programming (GEP) technique to build an objective function that represents the performance of a job and the correlation among the configuration parameters. For the purpose of optimization, Particle Swarm Optimization (PSO) is employed to find automatically an optimal or a near optimal configuration settings. The performance of the proposed work is intensively evaluated on a Hadoop cluster and the experimental results show that the proposed work enhances the performance of Hadoop significantly compared with the default settings.

Due to the imperative need to reduce the management costs, operators multiplex several concurrent applications in large datacenters. However, uncontrolled resource sharing between co-hosted applications often results in performance degradation problems, thus creating violations of service level agreements (SLAs) for service providers. Therefore, in order to meet per-application SLAs, per-application performance modeling for dynamic resource allocation in shared resource environments has recently become promising. We introduce Chorus, an interactive performance modeling framework for building application performance models incrementally and on the fly. It can be used to support complex, multi-tier resource allocation, and/or what-if performance inquiry in modern datacenters, such as Clouds. Chorus consists of (i) a declarative high-level language for providing semantic model guidelines, such as model templates, model functions, or sampling guidelines, from a sysadmin or a performance analyst, as model approximations to be learned or refined experimentally, (ii) a runtime engine for iteratively collecting experimental performance samples, validating and refining performance models. Chorus efficiently builds accurate models online, reuses and adjusts archival models over time, and combines them into an ensemble of models. We perform an experimental evaluation on a multi-tier server platform, using several industry- standard benchmarks. Our results show that Chorus is a flexible modeling framework and knowledge base for validating, extending and reusing existing models while adapting to new situations.

Due to the imperative need to reduce the management costs, operators multiplex several concurrent applications in large datacenters. However, uncontrolled resource sharing between co-hosted applications often results in performance degradation problems, thus creating violations of service level agreements (SLAs) for service providers. Therefore, in order to meet per-application SLAs, per-application performance modeling for dynamic resource allocation in shared resource environments has recently become promising. We introduce Chorus, an interactive performance modeling framework for building application performance models incrementally and on the fly. It can be used to support complex, multi-tier resource allocation, and/or what-if performance inquiry in modern datacenters, such as Clouds. Chorus consists of (i) a declarative high-level language for providing semantic model guidelines, such as model templates, model functions, or sampling guidelines, from a sysadmin or a performance analyst, as model approximations to be learned or refined experimentally, (ii) a runtime engine for iteratively collecting experimental performance samples, validating and refining performance models. Chorus efficiently builds accurate models online, reuses and adjusts archival models over time, and combines them into an ensemble of models. We perform an experimental evaluation on a multi-tier server platform, using several industry- standard benchmarks. Our results show that Chorus is a flexible modeling framework and knowledge base for validating, extending and reusing existing models while adapting to new situations.

Cloud computing provides an easy access to computing resources. Customers can acquire and release resources any time. However, it is not trivial to determine when and how many resources to allocate. Many applications running in the cloud face workload changes that affect their resource demand. The first thought is to plan capacity either for the average load or for the peak load. In the first case there is less cost incurred, but performance will be affected if the peak load occurs. The second case leads to money wastage, since resources will remain underutilized most of the time. Therefore there is a need for a more sophisticated resource provisioning techniques that can automatically scale the application resources according to workload demand and performance constrains. Large cloud providers such as Amazon, Microsoft, RightScale provide auto-scaling services. However, without the proper configuration and testing such services can do more harm than good. In this work I investigate application specific online resource allocation techniques that allow to dynamically adapt to incoming workload, minimize the cost of virtual resources and meet user-specified performance objectives.

Cloud computing provides an easy access to computing resources. Customers can acquire and release resources any time. However, it is not trivial to determine when and how many resources to allocate. Many applications running in the cloud face workload changes that affect their resource demand. The first thought is to plan capacity either for the average load or for the peak load. In the first case there is less cost incurred, but performance will be affected if the peak load occurs. The second case leads to money wastage, since resources will remain underutilized most of the time. Therefore there is a need for a more sophisticated resource provisioning techniques that can automatically scale the application resources according to workload demand and performance constrains. Large cloud providers such as Amazon, Microsoft, RightScale provide auto-scaling services. However, without the proper configuration and testing such services can do more harm than good. In this work I investigate application specific online resource allocation techniques that allow to dynamically adapt to incoming workload, minimize the cost of virtual resources and meet user-specified performance objectives.

With the widespread use of the Internet, the popularity of web applications has significantly increased. Such applications are subject to unpredictable workload conditions that vary from time to time. For example, an e-commerce website may face higher workloads than normal during festivals or promotional schemes. Such applications are critical and performance related issues, or service disruption can result in financial losses. Cloud computing with its attractive feature of dynamic resource provisioning (elasticity) is a perfect match to host such applications. The rapid growth in the usage of cloud computing model, as well as the rise in complexity of the web applications poses new challenges regarding the effective monitoring and management of the underlying cloud computational resources. This thesis investigates the state-of-the-art elastic methods including the models and techniques for the dynamic management and provisioning of cloud resources from a service provider perspective. An elastic controller is responsible to determine the optimal number of cloud resources, required at a particular time to achieve the desired performance demands. Researchers and practitioners have proposed many elastic controllers using versatile techniques ranging from simple if-then-else based rules to sophisticated optimisation, control theory and machine learning based methods. However, despite an extensive range of existing elasticity research, the aim of implementing an efficient scaling technique that satisfies the actual demands is still a challenge to achieve. There exist many issues that have not received much attention from a holistic point of view. Some of these issues include: 1) the lack of adaptability and static scaling behaviour whilst considering completely fixed approaches; 2) the burden of additional computational overhead, the inability to cope with the sudden changes in the workload behaviour and the preference of adaptability over reliability at runtime whilst considering the fully dynamic approaches; and 3) the lack of considering uncertainty aspects while designing auto-scaling solutions. This thesis seeks solutions to address these issues altogether using an integrated approach. Moreover, this thesis aims at the provision of qualitative elasticity rules. This thesis proposes a novel biologically-inspired switched feedback control methodology to address the horizontal elasticity problem. The switched methodology utilises multiple controllers simultaneously, whereas the selection of a suitable controller is realised using an intelligent switching mechanism. Each controller itself depicts a different elasticity policy that can be designed using the principles of fixed gain feedback controller approach. The switching mechanism is implemented using a fuzzy system that determines a suitable controller/- policy at runtime based on the current behaviour of the system. Furthermore, to improve the possibility of bumpless transitions and to avoid the oscillatory behaviour, which is a problem commonly associated with switching based control methodologies, this thesis proposes an alternative soft switching approach. This soft switching approach incorporates a biologically-inspired Basal Ganglia based computational model of action selection. In addition, this thesis formulates the problem of designing the membership functions of the switching mechanism as a multi-objective optimisation problem. The key purpose behind this formulation is to obtain the near optimal (or to fine tune) parameter settings for the membership functions of the fuzzy control system in the absence of domain experts’ knowledge. This problem is addressed by using two different techniques including the commonly used Genetic Algorithm and an alternative less known economic approach called the Taguchi method. Lastly, we identify seven different kinds of real workload patterns, each of which reflects a different set of applications. Six real and one synthetic HTTP traces, one for each pattern, are further identified and utilised to evaluate the performance of the proposed methods against the state-of-the-art approaches.

No description available.