Diabetes has become one of the most severe deceases due to an increasing number of diabetes patients globally. A large amount of digital data on diabetes has been collected through various channels. How to utilize these data sets to help doctors to make a decision on diagnosis, treatment and prediction of diabetic patients poses many challenges to the research community. The thesis investigates mathematical models with a focus on neural networks for large scale diabetes data modelling and analysis by utilizing modern computing technologies such as grid computing and cloud computing. These computing technologies provide users with an inexpensive way to have access to extensive computing resources over the Internet for solving data and computationally intensive problems. This thesis evaluates the performance of seven representative machine learning techniques in classification of diabetes data and the results show that neural network produces the best accuracy in classification but incurs high overhead in data training. As a result, the thesis develops MRNN, a parallel neural network model based on the MapReduce programming model which has become an enabling technology in support of data intensive applications in the clouds. By partitioning the diabetic data set into a number of equally sized data blocks, the workload in training is distributed among a number of computing nodes for speedup in data training. MRNN is first evaluated in small scale experimental environments using 12 mappers and subsequently is evaluated in large scale simulated environments using up to 1000 mappers. Both the experimental and simulations results have shown the effectiveness of MRNN in classification, and its high scalability in data training. MapReduce does not have a sophisticated job scheduling scheme for heterogonous computing environments in which the computing nodes may have varied computing capabilities. For this purpose, this thesis develops a load balancing scheme based on genetic algorithms with an aim to balance the training workload among heterogeneous computing nodes. The nodes with more computing capacities will receive more MapReduce jobs for execution. Divisible load theory is employed to guide the evolutionary process of the genetic algorithm with an aim to achieve fast convergence. The proposed load balancing scheme is evaluated in large scale simulated MapReduce environments with varied levels of heterogeneity using different sizes of data sets. All the results show that the genetic algorithm based load balancing scheme significantly reduce the makespan in job execution in comparison with the time consumed without load balancing.

RAID-6 is widely used to tolerate concurrent failures of any two disks to provide a higher level of reliability with the support of erasure codes. Among many implementations, one class of codes called Maximum Distance Separable (MDS) codes aims to offer data protection against disk failures with optimal storage efficiency. Typical MDS codes contain horizontal and vertical codes. However, because of the limitation of horizontal parity or diagonal/anti-diagonal parities used in MDS codes, existing RAID-6 systems suffer several important problems on performance and scalability, such as low write performance, unbalanced I/O, and high migration cost in the scaling process. To address these problems, in this dissertation, we design techniques for high performance and scalable RAID-6 systems. It includes high performance and load balancing erasure codes (H-Code and HDP Code), and Stripe-based Data Migration (SDM) scheme. We also propose a flexible MDS Scaling Framework (MDS-Frame), which can integrate H-Code, HDP Code and SDM scheme together. Detailed evaluation results are also given in this dissertation.

The purpose of this thesis was to analyze the problem of resource load balancing in virtualization clusters. Another aim was to implement a pilot version of resource load balancer for the VMware vSphere Standard-based virtualization cluster. The thesis also inspected available commercial and open source resource load balancers and examined their usability and effectiveness. While designing the custom solution, a modification of the greedy algorithm has been chosen to be used to determine which virtual machines should be migrated and to select their target hosts. Furthermore, experiments have been conducted to determine some parameters for the algorithm. Finally, it was experimentally verified that the implemented solution can be applied to effectively balance virtualization server workloads by live migrating virtual machines running on these hosts.

Shared resource contention is a significant problem in multi-core systems and can have a negative impact on the system. Memory contention occurs when the different cores in a processor access the same memory resource, resulting in a conflict. It is possible to limit memory contention through resource reservation where a part of the system or an application is reserved a partition of the shared resource. We investigated how applying memory bandwidth restrictions using MemGuard can aid in synchronizing execution times of parallel tasks. We further investigate when memory bandwidth restrictions are applicable. We conduct three experiments to investigate when bandwidth restrictions are applicable. Firstly, we conducted an experiment to pinpoint when the memory bandwidth saturates a core. Secondly, we investigated our adaptive memory partitioning scheme performance against static and no partitioning. Finally, we tested how our adaptive partitioning scheme and static partitioning can isolate a workload against an interfering memory intensive workload running on a separate core. As the experiments only were conducted on one system, pinpointing the general point of contention was difficult, seeing that it can differ significantly from system to system. Through our experiments, we see that memory bandwidth partitioning has the ability to decrease the execution time of feature detection algorithms, which means that memory bandwidth partitioning potentially can help threads to reach their synchronization points simultaneously.

Este trabalho apresenta uma análise de comportamento de índices de carga relacionados ao uso e à atividade de memória. São descritos 11 índices que refletem direta ou indiretamente a carga de memória. Esses índices podem ser obtidos facilmente no sistema peracional GNU/Linux através do sistema de arquivos /proc. Uma ferramenta de monitoramento foi criada para facilitar a análise de comportamento, podendo também servir para fornecer informações de carga para políticas de escalonamento. Os valores de cada índice foram monitorados durante a execução de uma carga de trabalho composta por aplicações reais que utilizam altas quantidades de memória. A partir dos resultados é possível descobrir a utilidade de cada índice, indicando como eles podem ser usados para auxiliar políticas de escalonamento a avaliar a carga de memória de uma máquina. Métricas de avaliação de carga originárias da combinação de mais de um índice são descritas. Essas métricas foram criadas em casos em que a combinação de mais de um índice permitia representar a carga de memória com mais precisão do que índices usados isoladamente. As métricas e índices pesquisados proporcionam formas precisas de representar a carga de memória em vários níveis, desde níveis de baixa utilização até níveis de saturação da memória principal com sintomas de ocorrência de thrashing. / This work presents an analysis of the behavior of load indices related to memory usage and activity. It describes 11 indices which reflect memory load, directly or indirectly. These indices may be easily obtained on the GNU/Linux operating system using the /proc filesystem. A monitoring tool has been developed to make the behavior analysis easier, but it can also be employed to provide load information to scheduling policies. The values of each index have been monitored during the execution of a workload composed by real memory-intensive applications. From the monitoring results, it is possible to find out the usefulness of each index, pointing out how it may be used to help scheduling policies evaluate the memory load of a certain computer. Load evaluation metrics have been created combining more than one index, with the objective of characterizing memory load more precisely than just one isolated index. The explored metrics and indices have shown to provide precise forms of representing memory load at several levels, from low utilization levels up to excessive main memory usage with symptoms of thrashing.

Este trabalho analisa o impacto causado no desempenho de uma simulação distribuída quando técnicas de particionamento convencionais são empregadas. Essas técnicas não levam em conta informações inerentes ao estado da simulação. Pelo fato da execução de uma simulação também estar sujeita a sofrer interferências da plataforma, informações sobre a potência computacional de cada recurso utilizado e sobre o tipo de simulação, podem ser aplicadas em seu particionamento. Foram utilizadas informações estáticas, geradas através da avaliação da plataforma com benchmarks, e dinâmicas, obtidas através de índices de carga. Os resultados obtidos da utilização destas técnicas se mostram atrativos, principalmente quando o objetivo é a execução das simulações em ambientes que não disponibilizam políticas de escalonamento específicas e sim políticas convencionais. Nos estudos de casos avaliados, observaram-se ganhos satisfatórios, como a redução de até 24% do tempo de execução, um aumento de até 22% de eficiência e 79% menos rollbacks causados. Percebe-se que dependendo do tempo que se dispõe e dos objetivos pretendidos, as técnicas convencionais podem ser empregadas em simulações distribuídas. Este trabalho também contribui com o aperfeiçoamento das duas ferramentas utilizadas: WARPED e AMIGO. Uma interface de comunicação entre as duas ferramentas foi desenvolvida, ampliando assim seus campos de utilização. / This work analyses the impact caused on distributed simulation performance when conventional partitioning techniques are employed. These techniques do not take into account inherent information about the state of the simulation. Since a simulation execution is subject to platform interference, information about the type of simulations and about the computational power of resources may be applied to the partitioning process. Static performance information, generated from evaluating the platform with benchmarks has been employed, as well as dynamic load information provided by load indices. The results obtained with this approach are attractive, mainly when the objective is to execute simulations on environments which make conventional scheduling policies available, instead of specific policies. The evaluated case studies show satisfactory performance gains of up to 24% of reduction in execution time, 22% of improvement in efficiency and reduction of up to 79% in rollbacks. So, depending on the available time and the aimed objectives, it is worth using conventional techniques to assist distributed simulation partitioning. This work also contributes to the improvement of both tools used in it: Warped and AMIGO. A communication interface has been developed to integrate the tools, extending their capabilities.

Algoritmos de balanceamento de carga são utilizados em sistemas distribuídos para homogeneizar a ocupação dos recursos computacionais disponíveis. A homogeneidade na ocupação do ambiente permite otimizar a alocação de recursos e, conseqüentemente, aumentar o desempenho das aplicações. Com o advento dos sistemas distribuídos de alta escala, fazem-se necessárias pesquisas para a construção de algoritmos de balanceamento de carga que sejam capazes de gerir com eficiência esses sistemas. Essa eficiência é medida através do número de mensagens geradas no ambiente, do suporte a ambientes heterogêneos, do uso de políticas que consomem poucos recursos do sistema, da estabilidade em alta carga, da escalabilidade do sistema e dos baixos tempos de resposta. Com o objetivo de atender as necessidades dos sistemas distribuídos de alta escala, este doutorado propõe, apresenta e avalia um novo algoritmo de balanceamento de carga denominado TLBA (Tree Load Balancing Algorithm). Esse algoritmo organiza os computadores do sistema em uma topologia lógica na forma de árvore, sobre a qual são executadas operações de balanceamento de carga. Para validar o TLBA foi construído um simulador que, submetido a testes, permitiu comprovar suas contribuições, que incluem: o baixo número de mensagens geradas pelas operações de balanceamento de carga; a estabilidade em altas cargas; os baixos tempos médios de resposta de processos. Para validar os resultados de simulação, foi construído um protótipo do TLBA. Esse protótipo confirmou os resultados de simulação e, conseqüentemente, as contribuições do algoritmo. / Load balancing algorithms are applied in distributed systems to homogenize the occupation of the available computational resources. The homogeneity of the environment occupation allows optimising the resource allocation and consequently, increasing the application performance. With the advent of the large-scale distributed systems, it was necessary to start researching the construction of load balancing algorithms which are able to manage these systems with efficiency. This efficiency is measured through the number of messages generated on the environment; the support to heterogeneous environments and the load balance policies which should spend the minimal resources time; the stability in overloaded situations; the system scalability; and the processes average response times, that should be small. With the aim to achieve the large-scale distributed systems requirements, this Ph.D. proposes, presents and evaluates a new load balancing algorithm named TLBA (Tree Load Balancing Algorithm). This algorithm arranges the computers on a logical network topology with a tree format. The load balancing operations are executed over this tree. To evaluate the TLBA algorithm, a simulator was built that was submitted to tests that confirmed the following characteristics: the small number of messages generated by the load balancing operations; the stability in overloaded situations; the small average processes response times. To validate the simulation results a TLBA prototype was implemented. This prototype confirmed the simulation results and consequently the contributions of the proposed algorithm.

" With the increase in traffic on the web, popular web sites get a large number of requests. Servers at these sites are sometimes unable to handle the large number of requests and clients to such sites experience long delays. One approach to overcome this problem is the distribution or replication of content over multiple servers. This approach allows for client requests to be distributed to multiple servers. Several techniques have been suggested to direct client requests to multiple servers. We discuss these techniques. With this work we hope to study the extent and method of content replication and distribution at web sites. To understand the distribution and replication of content we ran client programs to retrieve headers and bodies of web pages and observed the changes in them over multiple requests. We also hope to understand possible problems that could face clients to such sites due to caching and standardization of newer protocols like HTTP/1.1. The main contribution of this work is to understand the actual implementation of replicated and distributed content on multiple servers and its implication for clients. Our investigations showed issues with replicated and distributed content and its effects on caching due to incorrect identifers being send by different servers serving the same content. We were able to identify web sites doing application layer switching mechanisms like DNS and HTTP redirection. Lower layers of switching needed investigation of the HTTP responses from servers, which were hampered by insuffcient tags send by servers. We find web sites employ a large amount of distribution of embedded content and its ramifcations on HTTP/1.1 need further investigation. "

PANTS (PANTS Application Node Transparency System) is a suite of programs designed to add transparent load balancing to a Beowulf cluster so that processes are transfered among the nodes of the cluster to improve performance. PANTS provides the option of using one of several different load balancing policies, each having a different approach. This paper studies the scalability and performance of these policies on large clusters and under various workloads. We measure the performance of our policies on our current cluster, and use that performance data to build simulations to test the performance of the policies in larger clusters and under differing workloads. Two policies, one deterministic and one non-deterministic, are presented which offer optimal steady-state performance. We also present best practices and discuss the major challenges of load balancing policy design.

Submitted by Nara Lays Domingues Viana Oliveira (naradv) on 2015-07-16T15:53:35Z No. of bitstreams: 1 LUCAS.pdf: 1303250 bytes, checksum: 9f822df40e01702f7d665e6b6a631bf2 (MD5) / Made available in DSpace on 2015-07-16T15:53:35Z (GMT). No. of bitstreams: 1 LUCAS.pdf: 1303250 bytes, checksum: 9f822df40e01702f7d665e6b6a631bf2 (MD5) Previous issue date: 2012 / Nenhuma / Migração de processos é um pertinente mecanismo para oferecer balanceamento dinâmico de carga, principalmente em ambientes dinâmicos e heterogêneos. Em especial, esse tópico é importante para aplicações BSP (Bulk Synchronous Parallel) uma vez que elas compreendem execuções em fases, onde o tempo de cada superetapa é determinado pelo processo mais lento. Nesse contexto, esse trabalho apresenta o sistema jMigBSP. Ele permite a escrita de aplicações BSP em Java e seu diferencial diz respeito às facilidades de reescalonamento de objetos em duas maneiras: (i) usando diretivas de migração no código da aplicação e; (ii) através do balanceamento de carga automático em nível de middleware. Além das abordagens de reescalonamento, jMigBSP facilita a interação entre os objetos através de métodos para comunicação assíncrona e one-sided. O desenvolvimento de jMigBSP foi guiado pelas ideias de eficiência e flexibilidade. Em primeiro lugar, a eficiência é marcada pela preocupação com o seu desempenho se comparado com linguagens compiladas, bem como no próprio algoritmo de reescalonamento. Além disso, a flexibilidade está presente no tratamento do reescalonamento automático de objetos. A avaliação de jMigBSP compreendeu o desenvolvimento e a execução de duas aplicações BSP em um ambiente multicluster: (i) transformada rápida de Fourier e; (ii) compressão de imagens. Duas heurísticas para a seleção dos objetos candidatos à migração foram aplicadas na avaliação. A primeira seleciona um objeto BSP com o maior valor de PM (Potencial de Migração). A segunda escolhe uma percentagem de objetos baseado no maior PM. Os resultados mostram que jMigBSP oferece a oportunidade de ganhos de desempenho sem alterações no código da aplicação. jMigBSP torna possível ganhos de desempenho na casa de 29%, bem como produz uma baixa sobrecarga quando comparado com uma biblioteca de código nativo. Além disso, uma sobrecarga média de 5,52% foi observada no algoritmo de reescalonamento. Em geral, os resultados obtidos mostram na prática a teoria da migração de processos, onde aplicações computacionalmente intensivas (CPU-bound) são mais beneficiadas com a transferência de entidades (processos, tarefas, objetos etc.) para processadores mais rápidos. Considerando que a seleção de uma percentagem de objetos para migração se mostrou uma heurística eficiente, trabalhos futuros compreendem o desenvolvimento de outras que selecionam uma coleção de objetos sem a necessidade de parâmetros particulares para o reescalonador. / Process migration is an useful mechanism for runtime load balancing, mainly in heterogeneous and dynamic environments. In particular, this technique is important for Bulk Synchronous Parallel (BSP) applications. This kind of application is based in rounds, or supersteps, where the time of each superstep is determined by the slowest process. In this context, this work presents the jMigBSP system. It was designed to act over BSP-based Java applications and its differential approach concerns the offering of the rescheduling facility in two ways: (i) by using migration directives in the application code directly and; (ii) through automatic load balancing at middleware level. In addition, the presented library makes the object interaction easier by providing one-sided asynchronous communication. The development of jMigBSP was guided by the following ideas: efficiency and flexibility. First of all, the efficiency topic involves the performance relation with compiled languages (native code), as well as the time spent in the rescheduling algorithm itself. Moreover, the flexibility is present in the treatment of automatic object rescheduling. The evaluation of jMigBSP comprised the development and execution of two BSP applications in a multicluster environment: (i) fast Fourier transform and; (ii) Fractal image compression. Two heuristics were used for selecting the candidate objects for migration in the evaluation. The first heuristic chooses the BSP object that presents the highest PM (Potential of Migration) value. The second heuristic selects a percentage of objects based on the highest PM value. The results showed that jMigBSP offers an opportunity to get performance in an effortless manner to the programmer since its does not need modifications in the application code. jMigBSP makes possible gains of performance up to 29% as well as produces a low overhead when compared with a C-based library. Furthermore, an average overhead of 5,52% was observed in the rescheduling algorithm. In general, the results demonstrate in practice the theory of process migration, where computationally intensive applications (CPU-bound) are most benefited by the entities transferring (processes, tasks or objects) to faster processors. Considering that the selection of a percentage of objects for migration showed an efficient heuristic, future work includes the development of new mechanisms that select a collection of objects without the need to setup particular parameters to the rescheduler.