The emerging general purpose graphics processing units (GPGPU) computing has tremendously speeded up a great variety of commercial and scientific applications. The GPUs have become prevalent accelerators in current high performance clusters. Though the computational capacity per Watt of the GPUs is much higher than that of the CPUs, the hybrid GPU clusters still consume enormous power. To conserve energy on this kind of clusters is of critical significance. In this thesis, we seek energy conservative computing on the GPU accelerated servers. We introduce our studies as follows. First, we dissect the GPU memory hierarchy due to the fact that most of the GPU applications are suffering from the GPU memory bottleneck. We find that the conventional CPU cache models cannot be applied on the modern GPU caches, and the microbenchmarks to study the conventional CPU cache become invalid for the GPU. We propose the GPU-specified microbenchmarks to examine the GPU memory structures and properties. Our benchmark results verify that the design goal of the GPU has transformed from pure computation performance to better energy efficiency. Second, we investigate the impact of dynamic voltage and frequency scaling (DVFS), a successful energy management technique for CPUs, on the GPU platforms. Our experimental results suggest that GPU DVFS is still promising in conserving energy, but the patterns to save energy strongly differ from those of the CPU. Besides, the effect of GPU DVFS depends on the individual application characteristics. Third, we derive the GPU DVFS power and performance models from our experimental results, based on which we find the optimal GPU voltage and frequency setting to minimize the energy consumption of a single GPU task. We then study the problem of scheduling multiple tasks on a hybrid CPU-GPU cluster to minimize the total energy consumption by GPU DVFS. We design an effective offline scheduling algorithm which can reduce the energy consumption significantly. At last, we combine the GPU DVFS and dynamic resource sleep (DRS), another energy management technique, to further conserve the energy, for the online task scheduling on hybrid clusters. Though the idle energy consumption increases significantly compared to the offline problem, our online scheduling algorithm still achieves more than 30% of energy conservation with appropriate runtime GPU DVFS readjustments.

A fundamental assumption behind most machine learning (ML) algorithms and analyses is the sequential execution. That is, any update to the ML model can be immediately applied and the new model is always available for the next algorithmic step. This basic assumption, however, can be costly to realize, when the computation is carried out across multiple machines, linked by commodity networks that are usually 104 times slower than the memory speed due to fundamental hardware limitations. As a result, concurrent ML computation in the distributed settings often needs to handle delayed updates and perform learning in the presence of staleness. This thesis characterizes learning with staleness from three directions: (1) We extend the theoretical analyses of a number of classical ML algorithms, including stochastic gradient descent, proximal gradient descent on non-convex problems, and Frank-Wolfe algorithms, to explicitly incorporate staleness into their convergence characterizations. (2)We conduct simulation and large-scale distributed experiments to study the empirical effects of staleness on ML algorithms under indeterministic executions. Our results reveal that staleness is a key parameter governing the convergence speed for all considered ML algorithms, with varied ramifications. (3) We design staleness-minimizing parameter server systems by optimizing synchronization methods to effectively reduce the runtime staleness. The proposed optimization of a bounded consistency model utilizes the additional network bandwidths to communicate updates eagerly, relieving users of the burden to tune the staleness level. By minimizing staleness at the framework level, our system stabilizes diverging optimization paths and substantially accelerates convergence across ML algorithms without any modification to the ML programs.

The purpose of this thesis is to investigate and test the usage of HTTP/2 in dynamic adaptive video streaming as well as to take a look into how it can be used to benefit prefetching algorithms used with branched video. With a series of experiments the performance gains of using HTTP/2 rather than the older standard HTTP/1.1 has been investigated. The results has shown no significant change to player quality and buffer occupancy when using HTTP/2, though our tests has shown in a slight decrease in overall playback quality when using HTTP/2. When using a linear prefetch of two fragments an average quality improvement of 4.59% has been shown, however, the result is inconclusive due to variations in average quality between different values for how many fragments to prefetch. Average buffer occupancy has shown promise with a maximum increase of 12.58%, when using linear prefetch with three fragments. The values for buffer occupancy gains are conclusive. Two implementations for non-linear prefetching has been made. The first one uses HTTP/2 server push to deliver fragments for prefetching and the second one uses client-side invoked HTTP requests to pull fragments from the server. Using HTTP/2 server push has shown in a decrease of 2.5% in average total load time while using client-side pulling has shown in a decrease of 34% in average total load time.

Detta examensarbete omfattar utveckling och test av en metod för distribution av data i ett multidatabassystem. Testen av metoden har gjorts genom att den har implementerats i en fallstudie. Fallstudien berör en organisation som har behov av att integrera tre av sina databassystem. Detta för att information ska kunna utbytas mellan dessa. Integrationen av databassystemen har gjorts på ett sådant sätt att de data som distribueras från dessa skickas till en central databas, med SQL Server som databashanterare. Det verktyg som har valts att användas i examensarbetet för att lösa uppgiften är Data Transformation Services, DTS, i SQL Server.

Den här rapporten undersöker om gränsen mellan Terminal Services och Citrix XenApp har börjat suddas ut eller om man ännu kan dra en tydlig linje mellan dessa. Detta för att avgöra om organisationer som idag kör Windows Server 2003 med Citrix XenApp enbart klarar sig med de nya funktionerna i Windows Server 2008 Terminal Services och ändå uppfylla de krav och behov en organisation har. Arbetet genomfördes rent teoretiskt. För att avgöra likheter och skillnader mellan applikationerna, så användes till största delen ett ramverk. För att kunna få en bild av hur applikationerna tillämpas rent praktiskt, användes tidigare testar samt analyser. Resultatet visar att Terminal Services tagit ett större steg inom marknaden för applikationsvirtualisering. Slutsatsen antyder att Windows Server 2008 Terminal Services är lämplig för organisationer utan komplexa behov.

The purpose of this thesis is to develop software to stream input and output audio from a large number of computers in a network to one specific computer in the same network. This computer will save the audio to disk. The audio that is to be saved will consist mostly of spoken communication. The saved audio is to be used in a framework for modeling and visualization. There are three major problems involved in designing a software to fill this purpose: recording both input and output audio at the same time, efficiently receiving multiple audio-streams at once and designing an interface where finding and organizing the computers to record audio from is easy. The software developed to solve these problems consists of two parts; a server and a client. The server captures the input (microphone) and output (speaker) audio from a computer. To capture the output and input audio simultaneously an external application named Virtual Audio Cable (VAC) is used. The client connects to multiple servers and receives the captured audio. Each one of the client’s server-connections is handled by its own thread. To make it easy to find available servers an Automatic Server Discovery System has been developed. To simplify the organization of the servers they are displayed in a tree-view specifically designed for this purpose.

A World Wide Web é um meio de comunicação em constante crescimento, agregando diversos componentes e serviços em um ritmo acelerado. Os novos tipos de sites, tais como, o comércio eletrônico (e-commerce), notí?cia/informação (Web-publishing), vídeo sob demanda exigem ainda mais recursos do servidor. Nesse contexto, visando adequar a avaliação de desempenho aos novos ambientes da Web, o presente trabalho apresenta um estudo caracterizando diversos traces de servidores Web Apache, permitindo coletar dados importantes que definem a forma como os usuários e servidores interagem. Com esses dados, quatro tipos de categorias de sites foram analisados: Padrão (composto da média de todos os traces analisados), Acadêmico, Notícia/Informação e Tradicional. Nessa análise avaliam-se quatro aspectos: o intervalo de chegada, o código de resposta, o tipo objeto e o tamanho do objeto e ao final, modelos matemáticos são propostos para representar essas características. Além disso, este trabalho também desenvolveu um gerador de cargas de trabalho sintéticas, o W4Gen(World Wide Web Workload Generator). Com uma interface gráfica amigável, ele permite aos seus usuários gerar novas cargas com base nos modelos matemáticos. Além disso, ele também permite modificar as características essenciais para simular novos tipos de cargas. Para validar os resultados deste trabalho, utilizou-se o modelo de servidor Web com diferenciação de serviços (SWDS), verificando o desempenho em situações de sobregarga / World Wide Web is a media in constant growth, joining several components and services in an accelerated evolution. The new kinds of sites, such as, E-commerce, Web-publishing and demand video still uses more servers resources. In this context, adapting the newWeb environment to performance evaluation, the present work accomplished a characterization study of several Apache Web servers traces, allowing collect important data that define the form as users and servers interact. With these data, four types of sites categories were analyzed: Default (composed of all trace), Academic, Web-publishing and Traditional. In this analysis, it was evaluated four aspects: the arrival time, the status code, the object class and the object size and at the end, mathematical models are proposed to represent those characteristics. Furthermore, a synthetic workload generator was also developed. With a graphical interface, the W4Gen (World WideWebWorkload Generator), as called, allows the users to generate new workloads based on mathematical models. Besides, it also allows to modify the four essential aspects preseted above to simulate new types of workloads. Finally, to validate the results, the Web server model with differentiated services was used, verifying the performance in overload situations

A experiência com autoria de material didático multimídia para propósitos educacionais mostra um grande problema: como prover uma maneira de tratar objetos multimídia de modo que usuários inexperientes (como professores) possam estar aptos a projetar e construir suas próprias apresentações? A criação de tais apresentações envolve fatores como armazenamento, entrega, busca e apresentação de material multimídia (vídeo em especial). Uma infra-estrutura básica que armazene e entregue eficientemente os dados de vídeo é necessária, porém, outro ponto importante é organizar esses dados armazenados no servidor de forma a facilitar seu acesso por parte dos usuários. Neste trabalho, isto é alcançado através do uso de um sistema interativo de recuperação e gerenciamento de informações projetado para facilitar o acesso a itens (ou parte deles) armazenados no servidor. A principal característica de tal sistema é o uso de uma base de metadados contendo os atributos dos vídeos armazenados no servidor. Buscas podem ser feitas por título, assunto, tamanho, autor, conteúdo ou, mais importante no caso de material didático, por cenas ou frames específicos. O sistema foi implementado segundo uma abordagem cliente/servidor utilizando a linguagem de programação JAVA. A comunicação entre clientes e servidores é realizada através do uso do Visibroker 3.0, que é uma ferramenta de programação para Objetos Distribuídos segundo o padrão CORBA. O acesso aos dados a partir da base de metadados é realizado através do uso de um driver PostgreSQL que segue a API JDBC. Para propósitos de avaliação do sistema um player foi construído utilizando a ferramenta Java Media Framework (JMF). Foi realizada uma análise para a verificação do impacto da utilização das tecnologias CORBA e JDBC no sistema. Foi detectado que a utilização da tecnologia JDBC impõe um atraso muito mais significante que a utilização da tecnologia CORBA. Outra conclusão é que a utilização de metadados provê uma melhor interatividade em buscas, permite economia de tempo durante o processo de edição e provê economia de espaço de armazenamento através do compartilhamento de objetos como vídeos, cenas e frames. / The experience with authoring multimedia material for educational purposes shows a major problem: how to provide an easy and efficient way to handle multimedia objects in a manner that non-expert users (namely school teachers) can be able to design and build their own presentations? The creation of this presentations involves factors like storage, delivery, search and presentation of multimedia material (video in special). A basic infra-structure that stores and efficiently deliver the video data is needed. However, another important point is the organization of these data stored into the server in a way to facilitate the access to them from the users. In the system wich is the subject of this work, this is achived through the use of an interactive information management and retrieval system designed to facilitate the access to items (or parts of the items) stored in the server. The main characteristic of the system is the use of a metadata base which contains attributes of the videos stored in the server. Searches can be made by title, subject, length, author, content or, most important in the didatic multimedia material case, by a specific scene or frame. The system was built with JAVA programming language in a client/server way. The communication between clients and servers is realized through the use of the Visibroker 3.0, which is a Distributed Objects programming tool according to the CORBA standard. The data access from the metadata base use a PostgreSQL driver which follows the JDBC API. For evaluation purposes a playback tool was built using Java Media Framework (JMF). An analisys was carried out to verify the impact of the utilization of CORBA and JDBC technologies in the system. It was detected that JDBC technology utilization imposes a much more significate delay than the CORBA technology utilization. Another conclusion is that metadata utilization provide better interactivity searches, making the editing process faster and save storage space through the sharing of objects like videos, scenes and frames.

Esta dissertação de mestrado aborda a introdução de reconhecimento de sessões http em um modelo de servidor web com serviços diferenciados (SWDS). Algumas técnicas foram desenvolvidas com o objetivo de produzir diferenciação de serviços junto com garantias de que novas sessões poderiam ser aceitas no sistema. Esses objetivos constituem requisitos essenciais na Internet atual, especialmente para aplicações web modernas. Um novo esquema para controle de admissão de sessões foi desenvolvido e introduzido no modelo SWDS, considerando dois mecanismos para aceitar novas sessões, com garantia de nalização. O mecanismo que estima a capacidade do sistema de aceitar novas sessões, baseado em um modelo de sessão construído dinamicamente a partir da carga do sistema, é destacado. A proposta global deste trabalho também considera um controle de admissão de requisições, baseado em sessões, onde a nova política de atendimento criada mantém o sistema livre de sobrecargas e oferece atendimento diferenciado para as sessões. As políticas de negociação desenvolvidas para o controle de admissão de requisições tiveram um papel importante neste trabalho, contribuindo para a priorização do atendimento das sessões. Os resultados obtidos mostram que os controles propostos constituem estruturas fundamentais para a estabilidade do desempenho do sistema, tanto quanto os mecanismos desenvolvidos têm grande importância no atendimento das sessões e, portanto de seus clientes, através de uma abordagem baseada em diferenciação. / This MSc dissertation approaches the introduction of the HTTP sessions recognition in a web server model with diferentiated services (SWDS). Some techniques have been developed aiming at issuing diferentiation of services together with guarantees that new sessions could be accepted in the system. These aims constitute essential requirements for the current Internet, especially for modern web applications. A new scheme for the admission control system has been developed and introduced in the SWDS model, considering two mechanisms for accepting new sessions, with guarantee of their nalizations. The mechanism that estimates the system capacity of acceptance of a new session, based on a session model built dynamically from system workload information, is highlighted. The global propose of this work also considers a request admission control, based on sessions, where the new attendance polices created keep the system free from overloads and over diferentiated attendance for the sessions. The negotiation polices developed for request admission control had an important place in this work, contributing for the session attendance prioritization. The results reached show that the controls proposed comprise fundamental structures for system performance stability, as well as the mechanisms developed have great importance in attending sessions and, therefore, their clients by means of a diferentiation-based approach.

Virtualization allows computer users to utilize their resources more efficiently and effectively. Operating system that runs on top of the Virtual Machine or Hypervisor is called guest OS. The Virtual Machine is an abstraction of the real physical machine. The main aim of this thesis work was to analyze different kinds of virtualization software packages and to investigate their advantages and disadvantages. In addition, we analyzed the performance of the virtual software packages with a real operating system in terms of web services. Web Servers play an important role on the Internet. The response time and throughput for a web server are different for different virtualization software packages and between a real host and a virtual host. In this thesis, we analyzed the web server performance on Linux. We compared the throughput for three different virtualization software packages (VMware, QEMU, and Virtual Box). The performance results clearly indicate that the real machine performance is better than the performance of the virtual machines. VMware has the better performance compared to other virtual software packages.