No description available.

With scant research to draw upon with respect to the maintenance of vertical scales over time, decisions around the creation and performance of vertical scales over time necessarily suffers due to the lack of information. Undetected item parameter drift (IPD) presents one of the greatest threats to scale maintenance within an item response theory (IRT) framework. There is also still an outstanding question as to the utility of the Rasch model as an underlying viable framework for establishing and maintaining vertical scales. Even so, this model is currently used for scaling many state assessment systems. Most criticisms of the Rasch model in this context have not involved simulation. And most have not acknowledged conditions in which the model may function sufficiently to justify its use in vertical scaling. To address these questions, vertical scales were created from real data using the Rasch and 3PL models. Ability estimates were then generated to simulate a second (Time 2) administration. These simulated data were placed onto the base vertical scales using a horizontal vertical scaling approach and a mean-mean transformation. To examine the effects of IPD on vertical scale maintenance, several conditions of IPD were simulated to occur within each set of linking items. In order to evaluate the viability of using the Rasch model within a vertical scaling context, data were generated and calibrated at Time 2 within each model (Rasch and 3PL) as well as across each model (Rasch data generataion/3PL calibration, and vice versa). Results pertaining the first question of the effect IPD has on vertical scale maintenance demonstrate that IPD has an effect directly related to percentage of drifting linking items, the magnitude of IPD exhibited, and the direction. With respect to the viability of using the Rasch model within a vertical scaling context, results suggest that the Rasch model is perfectly viable within a vertical scaling context in which the model is appropriate for the data. It is also clearly evident that where data involve varying discrimination and guessing, use of the Rasch model is inappropriate.

Developmental score scales represent the performance of students along a continuum, where as students learn more they move higher along that continuum. Unidimensional item response theory (UIRT) vertical scaling has become a commonly used method to create developmental score scales. Research has shown that UIRT vertical scaling methods can be inconsistent in estimating grade-to-grade growth, within-grade variability, and separation of grade distributions (effect size) of developmental score scale. In particular the finding of scale shrinkage (decreasing within-grade score variability as grade-level increases) has led to concerns about and criticism of IRT vertical scales. The causes of scale shrinkage have yet to be fully understood. Real test data and simulation studies have been unable to provide complete answers as to why IRT vertical scaling inconsistencies occur. Violations of assumptions have been a commonly cited potential cause for the inconsistent results. For this reason, this dissertation is an extensive investigation into how violations of the three assumptions of UIRT vertical scaling - local item dependence, unidimensionality, and similar reliability of grade level tests - affect estimated developmental score scales. Simulated tests were developed that purposefully violated a UIRT vertical scaling assumption. Three sets of simulated tests were created to test the effect of violating a single assumption. First, simulated tests were created with increasing, decreasing, low, medium, and high local item dependence. Second, multidimensional simulated tests were created by varying the correlation between dimensions. Third, simulated tests with dissimilar reliability were created by varying item parameters characteristics of the grade level tests. Multiple versions of twelve simulated tests were used to investigate UIRT vertical scaling assumption violations. The simulated tests were calibrated under the UIRT model to purposefully violate an assumption of UIRT vertical scaling. Each simulated test version was replicated for 1000 random examinee samples to assess the bias and standard error of estimated grade-to-grade-growth, within-grade-variability, and separation-of-grade-distributions (effect size) of the estimated developmental score scales. The results suggest that when UIRT vertical scaling assumptions are violated the resulting estimated developmental score scales contain standard error and bias. For this study, the magnitude of standard error was similar across all simulated tests regardless of the assumption violation. However, bias fluctuated as a result of different types and magnitudes of UIRT vertical scaling assumption violations. More local item dependence resulted in more grade-to-grade-growth and separation-of-grade-distributions bias. And local item dependence resulted in developmental score scales that displayed scale expansion. Multidimensionality resulted in more grade-to-grade-growth and separation-of-grade-distributions bias when the correlation between dimensions was smaller. Multidimensionality resulted in developmental score scales that displayed scale expansion. Dissimilar reliability of grade level tests resulted in more grade-to-grade-growth bias and minimal separation-of-grade-distributions bias. Dissimilar reliability of grade level tests resulted in scale expansion or scale shrinkage depending on the item characteristics of the test. Limitations of this study and future research are discussed.

Cloud computing is a constantly developing topic that reaches most of the people in the world on a daily basis. Almost every website and mobile application is hosted through a cloud provider. Two of the most important metrics for customers is performance and availability. Current tools that mea- sure availability are using the Internet Control Message Protocol (ICMP) to monitor availability, which has shown to be unreliable. This thesis suggests a new way of monitoring both availability and response time by using Hypertext Transfer Protocol (HTTP). Through HTTP, we are able to reach both the front-end of the cloud service (just as ICMP), but also deeper, to find failures in the back-end, that ICMP would miss. With our monitoring tool, we have monitored five different cloud data centers during one week. We found that cloud providers are not always keeping their promised SLA and it might be up to the cloud customers to reach a higher availability. We also perform load tests to analyze how vertical and horizontal scaling performs with regards to response time. Our analysis concludes that, at this time, vertical scaling outperforms horizontal scaling when it comes to response time. Even when this is the case, we suggest that developers should build applications that are horizontally scalable. With a horizontally scalable application and our monitoring tool combined, we can reach higher availability than is currently possible.

No description available.

Cloud computing’s growing popularity is based on the cloud’s flexibility and the availability of a huge amount of resources. Today, cloud providers offer a wide range of predefined solutions, VM (virtual machine) sizes and customization differing in performance, support and price. In this thesis it is investigated how to achieve cost minimization within specified performance goals for a commercial service with computation occurring in a repetitive pattern. A promising multilevel queue scheduling and a set of auto-scaling rules to fulfil computation deadlines and job prioritization and lower server cost is presented. In addition, an investigation to find an optimal VM size in the sense of cost and performance points out further areas of cloud service optimization.

Submitted by Lucienne Costa (lucienneferreira@ufcg.edu.br) on 2018-08-30T18:21:31Z No. of bitstreams: 1 IURY GREGORY MELO FERREIRA – DISSERTAÇÃO (PPGCC) 2017.pdf: 3496497 bytes, checksum: b497c83bd5c1b1ab2be30ab67272f5cd (MD5) / Made available in DSpace on 2018-08-30T18:21:31Z (GMT). No. of bitstreams: 1 IURY GREGORY MELO FERREIRA – DISSERTAÇÃO (PPGCC) 2017.pdf: 3496497 bytes, checksum: b497c83bd5c1b1ab2be30ab67272f5cd (MD5) Previous issue date: 2017-12-18 / Em ambientes de Computação na Nuvem, principalmente os que utilizam o modelo de infraestrutura como um serviço, a característica de elasticidade no provisionamento de recursos traz consigo a necessidade de gerenciar os recursos físicos de forma apropriada para preservar a qualidade de serviço aos seus usuários, e o bom desempenho da infraestrutura. Este trabalho propõe heurísticas que são capazes de auxiliar no balanceamento de carga dos servidores em uma infraestrutura de nuvem, propondo migrações para diminuir a sobrecarga nos servidores que foram identificados como sobrecarregados,visto que, como passar do tempo há uma variação natural na quantidade de recursos em uso. Esta variação é uma consequência da remoção ou adição de aplicações, ou até mesmo de tentativas de melhoramento do desempenho das aplicações através do provisionamento vertical. Uma ferramenta foi implementada para fazer uso dos algoritmos das heurísticas e assim auxiliar nos experimentos para a validação das mesmas. As métricas utilizadas vem diretamente de servidores heterogêneos da nuvem OpenStack do Laboratório de Sistemas Distribuídos. Os resultados obtidos mostram que além da diminuição no consumo de CPU dos servidores dos quais que estavam sobrecarregados, também é possível melhorar o desempenho destes servidores em alguns casos. / In CloudComputingenvironments,especiallythoseusingtheinfrastructureasaservice model, theelasticitycharacteristicinresourceprovisioningcomeswiththeneedtomanage resources sothequalityofservicecancontinuetobeguaranteedtousersandalsoto maintain agoodperformanceoftheinfrastructure.Thisworkproposesheuristicsthat are abletoassistintheloadbalancingoftheserversinaCloudinfrastructure,proposing migrations toreducetheoverheadintheserversthatwereidentifiedasoverloaded,since with thepassageoftimethereisanaturalvariationintheamountofresourcesinuse.This variationinaconsequenceofremovaloradditionofapplicationsandevenoftheusageof verticalscalingtoimproveapplication’sperformance.Atoolwasimplementedtomake use oftheheuristicalgorithmsandthustoaidintheexperimentsandtheirvalidation,the metrics usedcomedirectlyfromheterogeneousserversoftheOpenStackCloudofthe DistributedSystemsLaboratory.TheresultsshowthatinadditiontothedecreaseinCPU consumption ofserversthatwereoverloaded,itisalsopossibletoimprovetheperformance of theseserversinsomecases.

Cloud computing and containerization has been on the rise in recent years and have become important areas of research and development in the field of computer science. One of the challenges in distributed and cloud computing is to predict the resource utilization of the nodes that run the applications and services. This is especially relevant for container-based platforms such as Kubernetes. Predicting the resource utilization of a Kubernetes cluster can help optimize the performance, reliability, and cost-effectiveness of the platform. This thesis focuses on how well different resources in a cluster can be predicted using machine learning techniques. The approach consists of 3 main steps: data collection, data extraction and pre-processing, and data analysis. The data collection step involves stressing the system with a load-generator called Locust and collecting data from Locust and collecting data from Kubernetes with the use of Prometheus. The data pre-processing and extraction step involves extracting relevant data and transforming it into a suitable format for the machine learning models. The final step involves applying different machine learning models to the data and evaluating their accuracy. The results of this thesis illustrate that machine learning can work well for predicting resources in a cluster based on how stressed the system is and that the best performing machine learning model tested was Support Vector Machine with a polynomial kernel. / Cloud computing och containerisering har ökat de senaste åren och har blivit viktiga områden för forskning och utveckling inom datavetenskap. En av utmaningarna inom distribuerad och cloud computing är att förutsäga resursutnyttjandet av de noder som kör applikationerna och tjänsterna. Detta är särskilt relevant för containerbaserade plattformar som Kubernetes. Att förutsäga resursutnyttjandet av ett Kubernetes-kluster kan hjälpa med att optimera plattformens prestanda, tillförlitlighet och kostnadseffektivitet. Denna avhandling fokuserar på hur väl olika resurser i ett kluster kan förutsägas med hjälp av maskininlärningstekniker. Tillvägagångssättet består av 3 huvudsteg: datainsamling, dataextraktion och för-processering, samt dataanalys. Datainsamlingssteget innebär att stressa systemet med en load-generator som heter Locust och samla in data från Locust och även samla in data från Kubernetes med hjälp av Prometheus. Steget för för-processering och extrahering av data innefattar att extrahera relevant data och omvandla den till ett lämpligt format för maskininlärningsmodellerna. Det sista steget innefattar att tillämpa olika maskininlärningsmodeller på data och utvärdera deras noggrannhet. Resultaten av denna avhandling demonstrerar att maskininlärning kan fungera bra för att förutsäga resurser i ett kluster baserat på hur stressat systemet är och att den bäst presterande maskininlärningsmodellen som testades var Support Vector Machine med en polynom-kernel.

Guidelines to screen and select common items for vertical scaling have been adopted from equating. Differences between vertical scaling and equating suggest that these guidelines may not apply to vertical scaling in the same way that they apply to equating. For example, in equating the examinee groups are assumed to be randomly equivalent, but in vertical scaling the examinee groups are assumed to possess different levels of proficiency. Equating studies that examined the characteristics of the common-item set stress the importance of careful item selection, particularly when groups differ in ability level. Since in vertical scaling cross-level ability differences are expected, the common items' psychometric characteristics become even more important in order to obtain a correct interpretation of students' academic growth. This dissertation applied two screening criteria and two selection approaches to investigate how changes in the composition of the linking sets impacted the nature of students' growth when creating vertical scales for two elementary mathematics tests. The purpose was to observe how well these equating guidelines were applied in the context of vertical scaling. Two separate datasets were analyzed to observe the impact of manipulating the common items' content area and targeted curricular grade level. The same Rasch scaling method was applied for all variations of the linking set. Both the robust z procedure and a variant of the 0.3-logit difference procedure were used to screen unstable common items from the linking sets. (In vertical scaling, a directional item-difficulty difference must be computed for the 0.3-logit difference procedure.) Different combinations of stable common items were selected to make up the linking sets. The mean/mean method was used to compute the equating constant and linearly transform the students' test scores onto the base scale. A total of 36 vertical scales were created. The results indicated that, although the robust z procedure was a more conservative approach to flagging unstable items, the robust z and the 0.3-logit difference procedure produced similar interpretations of students' growth. The results also suggested that the choice of grade-level-targeted common items affected the estimates of students' grade-to-grade growth, whereas the results regarding the choice of content-area-specific common items were inconsistent. The findings from the Geometry and Measurement dataset indicated that the choice of content-area-specific common items had an impact on the interpretation of students' growth, while the findings from the Algebra and Data Analysis/Probability dataset indicated that the choice of content-area-specific common items did not appear to significantly affect students' growth. A discussion of the limitations of the study and possible future research is presented.