A computação em grid (ou computação em grade) emergiu como uma área de pesquisa importante por permitir o compartilhamento de recursos computacionais geograficamente distribuídos entre vários usuários. Contudo, a heterogeneidade e a dinâmica do comportamento dos recursos em ambientes de grid tornam complexos o desenvolvimento e a execução de aplicações. OurGrid é uma plataforma de software que procura contornar estas dificuldades: além de permitir a execução de aplicações distribuídas em ambientes de computação em grid, oferece e gerencia um esquema de troca de favores entre usuários. Neste esquema, instituições (ou usuários) que possuam recursos ociosos podem oferecê-los a outros que deles necessitem. Quanto mais um domínio oferecer recursos ao grid, mais será favorecido quando precisar, ou seja, terá prioridade mais alta quando requisitar máquinas ao grid. O software MyGrid é o principal componente do OurGrid. É através dele que o usuário interage com o grid, submetendo e gerenciando suas aplicações. No modelo de execução do MyGrid, as tarefas são lançadas por um nó central que coordena todo o escalonamento de tarefas que serão executadas no grid. Este nó apresenta uma fragilidade caracterizada na literatura como "ponto único de falhas", pois seu colapso faz com que os resultados do processamento corrente sejam perdidos. Isto pode significar horas ou, até mesmo, dias de processamento perdido, dependendo das aplicações. Visando suprir esta deficiência, este trabalho descreve o funcionamento e a implementação de um mecanismo de checkpointing (ou salvamento de estado), usado como base para a recuperação por retorno, que permite ao sistema voltar a um estado consistente, minimizando a perda de dados, após uma falha no nó central do MyGrid. Assim, ele salva, de forma estável, o estado da aplicação (estruturas de dados e informações de controle imprescindíveis) capaz de restaurar o sistema após o colapso, oferecendo uma alternativa à sua característica de ponto único de falhas. Os checkpoints são obtidos e salvos a cada mudança de estado do escalonador de tarefas do nó central. A eficiência do mecanismo de recuperação é comprovada através de experimentos que exercitam este mecanismo em cenários com diferentes características, visando validar e avaliar o impacto real no desempenho do MyGrid. / The grid computing has emerged as an important research area because it allows sharing geographically distributed computing resources among several users. However, resources in a grid are highly heterogeneous and dynamic, turning complex the development and the execution of applications. OurGrid is a software platform that intends to reduce these difficulties. Besides allowing the execution of distributed applications in grid environments, it offers and gives support to an exchange of favors between users. In this way, institutions (or users) that have idle resources can offer them to other users. The more resources a domain offers to the grid, the more it will be favored when in need. It will have higher priority when requesting machines to grid. MyGrid software is the main component of OurGrid: it constitutes the interface for user interaction as well as application submission and management. In the execution model of MyGrid, tasks are launched by a central node (home-machine), which manages the scheduling of tasks to be executed in the grid. This node constitutes a "single point of failure", because its crash causes the loss of results of the previous processing. Depending on the particular applications, this loss can be the result of hours or days of processing time. This dissertation aims to reduce the consequences of this problem offering an alternative to the single point of failure: here is proposed and implemented a checkpointing mechanism, used as basis for the rollback recovery. Checkpoints are taken synchronously with the state changes of the scheduler on the central node. After a failure affecting the home-machine of MyGrid, the system recovers information on the state of the application (data structures and essential control information) and results of previous computation, saved in stable storage, minimizing the loss of data. The efficiency of the recovery mechanism and its impact over MyGrid are evaluated through experiments that exercise this mechanism in scenarios with different characteristics.

A computação em grid (ou computação em grade) emergiu como uma área de pesquisa importante por permitir o compartilhamento de recursos computacionais geograficamente distribuídos entre vários usuários. Contudo, a heterogeneidade e a dinâmica do comportamento dos recursos em ambientes de grid tornam complexos o desenvolvimento e a execução de aplicações. OurGrid é uma plataforma de software que procura contornar estas dificuldades: além de permitir a execução de aplicações distribuídas em ambientes de computação em grid, oferece e gerencia um esquema de troca de favores entre usuários. Neste esquema, instituições (ou usuários) que possuam recursos ociosos podem oferecê-los a outros que deles necessitem. Quanto mais um domínio oferecer recursos ao grid, mais será favorecido quando precisar, ou seja, terá prioridade mais alta quando requisitar máquinas ao grid. O software MyGrid é o principal componente do OurGrid. É através dele que o usuário interage com o grid, submetendo e gerenciando suas aplicações. No modelo de execução do MyGrid, as tarefas são lançadas por um nó central que coordena todo o escalonamento de tarefas que serão executadas no grid. Este nó apresenta uma fragilidade caracterizada na literatura como "ponto único de falhas", pois seu colapso faz com que os resultados do processamento corrente sejam perdidos. Isto pode significar horas ou, até mesmo, dias de processamento perdido, dependendo das aplicações. Visando suprir esta deficiência, este trabalho descreve o funcionamento e a implementação de um mecanismo de checkpointing (ou salvamento de estado), usado como base para a recuperação por retorno, que permite ao sistema voltar a um estado consistente, minimizando a perda de dados, após uma falha no nó central do MyGrid. Assim, ele salva, de forma estável, o estado da aplicação (estruturas de dados e informações de controle imprescindíveis) capaz de restaurar o sistema após o colapso, oferecendo uma alternativa à sua característica de ponto único de falhas. Os checkpoints são obtidos e salvos a cada mudança de estado do escalonador de tarefas do nó central. A eficiência do mecanismo de recuperação é comprovada através de experimentos que exercitam este mecanismo em cenários com diferentes características, visando validar e avaliar o impacto real no desempenho do MyGrid. / The grid computing has emerged as an important research area because it allows sharing geographically distributed computing resources among several users. However, resources in a grid are highly heterogeneous and dynamic, turning complex the development and the execution of applications. OurGrid is a software platform that intends to reduce these difficulties. Besides allowing the execution of distributed applications in grid environments, it offers and gives support to an exchange of favors between users. In this way, institutions (or users) that have idle resources can offer them to other users. The more resources a domain offers to the grid, the more it will be favored when in need. It will have higher priority when requesting machines to grid. MyGrid software is the main component of OurGrid: it constitutes the interface for user interaction as well as application submission and management. In the execution model of MyGrid, tasks are launched by a central node (home-machine), which manages the scheduling of tasks to be executed in the grid. This node constitutes a "single point of failure", because its crash causes the loss of results of the previous processing. Depending on the particular applications, this loss can be the result of hours or days of processing time. This dissertation aims to reduce the consequences of this problem offering an alternative to the single point of failure: here is proposed and implemented a checkpointing mechanism, used as basis for the rollback recovery. Checkpoints are taken synchronously with the state changes of the scheduler on the central node. After a failure affecting the home-machine of MyGrid, the system recovers information on the state of the application (data structures and essential control information) and results of previous computation, saved in stable storage, minimizing the loss of data. The efficiency of the recovery mechanism and its impact over MyGrid are evaluated through experiments that exercise this mechanism in scenarios with different characteristics.

abstract: Soft errors are considered as a key reliability challenge for sub-nano scale transistors. An ideal solution for such a challenge should ultimately eliminate the effect of soft errors from the microprocessor. While forward recovery techniques achieve fast recovery from errors by simply voting out the wrong values, they incur the overhead of three copies execution. Backward recovery techniques only need two copies of execution, but suffer from check-pointing overhead. In this work I explored the efficiency of integrating check-pointing into the application and the effectiveness of recovery that can be performed upon it. After evaluating the available fine-grained approaches to perform recovery, I am introducing InCheck, an in-application recovery scheme that can be integrated into instruction-duplication based techniques, thus providing a fast error recovery. The proposed technique makes light-weight checkpoints at the basic-block granularity, and uses them for recovery purposes. To evaluate the effectiveness of the proposed technique, 10,000 fault injection experiments were performed on different hardware components of a modern ARM in-order simulated processor. InCheck was able to recover from all detected errors by replaying about 20 instructions, however, the state of the art recovery scheme failed more than 200 times. / Dissertation/Thesis / Masters Thesis Computer Science 2016

Fault tolerance is due to the semiconductor technology development important, not only for safety-critical systems but also for general-purpose (non-safety critical) systems. However, instead of guaranteeing that deadlines always are met, it is for general-purpose systems important to minimize the average execution time (AET) while ensuring fault tolerance. For a given job and a soft (transient) no-error probability, we define mathematical formulas for AET using voting (active replication), rollback-recovery with checkpointing (RRC) and a combination of these (CRV) where bus communication overhead is included. And, for a given multi-processor system-on-chip (MPSoC), we define integer linear programming (ILP) models that minimize the AET including bus communication overhead when: (1) selecting the number of checkpoints when using RRC or a combination where RRC is included, (2) finding the number of processors and job-to-processor assignment when using voting or a combination where voting is used, and (3) defining fault tolerance scheme (voting, RRC or CRV) per job and defining its usage for each job. Experiments demonstrate significant savings in AET.

A computação em grid (ou computação em grade) emergiu como uma área de pesquisa importante por permitir o compartilhamento de recursos computacionais geograficamente distribuídos entre vários usuários. Contudo, a heterogeneidade e a dinâmica do comportamento dos recursos em ambientes de grid tornam complexos o desenvolvimento e a execução de aplicações. OurGrid é uma plataforma de software que procura contornar estas dificuldades: além de permitir a execução de aplicações distribuídas em ambientes de computação em grid, oferece e gerencia um esquema de troca de favores entre usuários. Neste esquema, instituições (ou usuários) que possuam recursos ociosos podem oferecê-los a outros que deles necessitem. Quanto mais um domínio oferecer recursos ao grid, mais será favorecido quando precisar, ou seja, terá prioridade mais alta quando requisitar máquinas ao grid. O software MyGrid é o principal componente do OurGrid. É através dele que o usuário interage com o grid, submetendo e gerenciando suas aplicações. No modelo de execução do MyGrid, as tarefas são lançadas por um nó central que coordena todo o escalonamento de tarefas que serão executadas no grid. Este nó apresenta uma fragilidade caracterizada na literatura como "ponto único de falhas", pois seu colapso faz com que os resultados do processamento corrente sejam perdidos. Isto pode significar horas ou, até mesmo, dias de processamento perdido, dependendo das aplicações. Visando suprir esta deficiência, este trabalho descreve o funcionamento e a implementação de um mecanismo de checkpointing (ou salvamento de estado), usado como base para a recuperação por retorno, que permite ao sistema voltar a um estado consistente, minimizando a perda de dados, após uma falha no nó central do MyGrid. Assim, ele salva, de forma estável, o estado da aplicação (estruturas de dados e informações de controle imprescindíveis) capaz de restaurar o sistema após o colapso, oferecendo uma alternativa à sua característica de ponto único de falhas. Os checkpoints são obtidos e salvos a cada mudança de estado do escalonador de tarefas do nó central. A eficiência do mecanismo de recuperação é comprovada através de experimentos que exercitam este mecanismo em cenários com diferentes características, visando validar e avaliar o impacto real no desempenho do MyGrid. / The grid computing has emerged as an important research area because it allows sharing geographically distributed computing resources among several users. However, resources in a grid are highly heterogeneous and dynamic, turning complex the development and the execution of applications. OurGrid is a software platform that intends to reduce these difficulties. Besides allowing the execution of distributed applications in grid environments, it offers and gives support to an exchange of favors between users. In this way, institutions (or users) that have idle resources can offer them to other users. The more resources a domain offers to the grid, the more it will be favored when in need. It will have higher priority when requesting machines to grid. MyGrid software is the main component of OurGrid: it constitutes the interface for user interaction as well as application submission and management. In the execution model of MyGrid, tasks are launched by a central node (home-machine), which manages the scheduling of tasks to be executed in the grid. This node constitutes a "single point of failure", because its crash causes the loss of results of the previous processing. Depending on the particular applications, this loss can be the result of hours or days of processing time. This dissertation aims to reduce the consequences of this problem offering an alternative to the single point of failure: here is proposed and implemented a checkpointing mechanism, used as basis for the rollback recovery. Checkpoints are taken synchronously with the state changes of the scheduler on the central node. After a failure affecting the home-machine of MyGrid, the system recovers information on the state of the application (data structures and essential control information) and results of previous computation, saved in stable storage, minimizing the loss of data. The efficiency of the recovery mechanism and its impact over MyGrid are evaluated through experiments that exercise this mechanism in scenarios with different characteristics.

Diplomarbeit im Schnittgebiet von Cluster- und Grid-Computing. Einbindung verteilter Ressourcen in eine Infrastruktur. Realisierung einer einfachen Checkpointing-Umgebung.

京都大学 / 0048 / 新制・課程博士 / 博士(情報学) / 甲第21280号 / 情博第674号 / 新制||情||116(附属図書館) / 京都大学大学院情報学研究科システム科学専攻 / (主査)教授 田中 利幸, 教授 山下 信雄, 准教授 増山 博之, 教授 笠原 正治 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

High availability is a critical feature of data centers, cloud, and cluster computing environments. Replication is a classical approach to increase service availability by providing redundancy. However, traditional replication methods are increasingly unattractive for deployment due to several limitations such as application-level non-transparency, non-isolation of applications (causing security vulnerabilities), complex system management, and high cost. Virtualization overcomes these limitations through another layer of abstraction, and provides high availability through virtual machine (VM) live migration: a guest VM image running on a primary host is transparently check-pointed and migrated, usually at a high frequency, to a backup host, without pausing the VM; the VM is resumed from the latest checkpoint on the backup when a failure occurs. A virtual cluster (VC) generalizes the VM concept for distributed applications and systems: a VC is a set of multiple VMs deployed on different physical machines connected by a virtual network. This dissertation presents a set of VM live migration techniques, their implementations in the Xen hypervisor and Linux operating system kernel, and experimental studies conducted using benchmarks (e.g., SPEC, NPB, Sysbench) and production applications (e.g., Apache webserver, SPECweb). We first present a technique for reducing VM migration downtimes called FGBI. FGBI reduces the dirty memory updates that must be migrated during each migration epoch by tracking memory at block granularity. Additionally, it determines memory blocks with identical content and shares them to reduce the increased memory overheads due to block-level tracking granularity, and uses a hybrid compression mechanism on the dirty blocks to reduce the migration traffic. We implement FGBI in the Xen hypervisor and conduct experimental studies, which reveal that the technique reduces the downtime by 77% and 45% over competitors including LLM and Remus, respectively, with a performance overhead of 13%. We then present a lightweight, globally consistent checkpointing mechanism for virtual cluster, called VPC, which checkpoints the VC for immediate restoration after (one or more) VM failures. VPC predicts the checkpoint-caused page faults during each checkpointing interval, in order to implement a lightweight checkpointing approach for the entire VC. Additionally, it uses a globally consistent checkpointing algorithm, which preserves the global consistency of the VMs' execution and communication states, and only saves the updated memory pages during each checkpointing interval. Our Xen-based implementation and experimental studies reveal that VPC reduces the solo VM downtime by as much as 45% and reduces the entire VC downtime by as much as 50% over competitors including VNsnap, with a memory overhead of 9% and performance overhead of 16%. The dissertation's third contribution is a VM resumption mechanism, called VMresume, which restores a VM from a (potentially large) checkpoint on slow-access storage in a fast and efficient way. VMresume predicts and preloads the memory pages that are most likely to be accessed after the VM's resumption, minimizing otherwise potential performance degradation due to cascading page faults that may occur on VM resumption. Our experimental studies reveal that VM resumption time is reduced by an average of 57% and VM's unusable time is reduced by 73.8% over native Xen's resumption mechanism. Traditional VM live migration mechanisms are based on hypervisors. However, hypervisors are increasingly becoming the source of several major security attacks and flaws. We present a mechanism called HSG-LM that does not involve the hypervisor during live migration. HSG-LM is implemented in the guest OS kernel so that the hypervisor is completely bypassed throughout the entire migration process. The mechanism exploits a hybrid strategy that reaps the benefits of both pre-copy and post-copy migration mechanisms, and uses a speculation mechanism that improves the efficiency of handling post-copy page faults. We modify the Linux kernel and develop a new page fault handler inside the guest OS to implement HSG-LM. Our experimental studies reveal that the technique reduces the downtime by as much as 55%, and reduces the total migration time by as much as 27% over competitors including Xen-based pre-copy, post-copy, and self-migration mechanisms. In a virtual cluster environment, one of the main challenges is to ensure equal utilization of all the available resources while avoiding overloading a subset of machines. We propose an efficient load balancing strategy using VM live migration, called DCbalance. Differently from previous work, DCbalance records the history of mappings to inform future placement decisions, and uses a workload-adaptive live migration algorithm to minimize VM downtime. We improve Xen's original live migration mechanism and implement the DCbalance technique, and conduct experimental studies. Our results reveal that DCbalance reduces the decision generating time by 79%, the downtime by 73%, and the total migration time by 38%, over competitors including the OSVD virtual machine load balancing mechanism and the DLB (Xen-based) dynamic load balancing algorithm. The dissertation's final contribution is a technique for VM live migration in Wide Area Networks (WANs), called FDM. In contrast to live migration in Local Area Networks (LANs), VM migration in WANs involve migrating disk data, besides memory state, because the source and the target machines do not share the same disk service. FDM is a fast and storage-adaptive migration mechanism that transmits both memory state and disk data with short downtime and total migration time. FDM uses page cache to identify data that is duplicated between memory and disk, so as to avoid transmitting the same data unnecessarily. We implement FDM in Xen, targeting different disk formats including raw and Qcow2. Our experimental studies reveal that FDM reduces the downtime by as much as 87%, and reduces the total migration time by as much as 58% over competitors including pre-copy or post-copy disk migration mechanisms and the disk migration mechanism implemented in BlobSeer, a widely used large-scale distributed storage service. / Ph. D.

Yes / Failure Prediction has long known to be a challenging problem. With the evolving trend of technology and growing complexity of high-performance cloud data centre infrastructure, focusing on failure becomes very vital particularly when designing systems for the next generation. The traditional runtime fault-tolerance (FT) techniques such as data replication and periodic check-pointing are not very effective to handle the current state of the art emerging computing systems. This has necessitated the urgent need for a robust system with an in-depth understanding of system and component failures as well as the ability to predict accurate potential future system failures. In this paper, we studied data in-production-faults recorded within a five years period from the National Energy Research Scientific computing centre (NERSC). Using the data collected from the Computer Failure Data Repository (CFDR), we developed an effective failure prediction model focusing on high-performance cloud data centre infrastructure. Using the Auto-Regressive Moving Average (ARMA), our model was able to predict potential future failures in the system. Our results also show a failure prediction accuracy of 95%, which is good.

Consider a time-dependent optimal control problem, where the state evolution is described by an initial value problem. There are a variety of numerical methods to solve these problems. The so-called indirect approach is considered detailed in this thesis. The indirect methods solve decoupled boundary value problems resulting from the necessary conditions for the optimal control problem. The so-called Pantoja method describes a computationally efficient stage-wise construction of the Newton direction for the discrete-time optimal control problem. There are many relationships between multiple shooting techniques and Pantoja method, which are investigated in this thesis. In this context, the equivalence of Pantoja method and multiple shooting method of Riccati type is shown. Moreover, Pantoja method is extended to the case where the state equations are discretised using one of implicit numerical methods. Furthermore, the concept of symplecticness and Hamiltonian systems is introduced. In this regard, a suitable numerical method is presented, which can be applied to unconstrained optimal control problems. It is proved that this method is a symplectic one. The iterative solution of optimal control problems in ordinary differential equations by Pantoja or Riccati equivalent methods leads to a succession of triple sweeps through the discretised time interval. The second (adjoint) sweep relies on information from the first (original) sweep, and the third (final) sweep depends on both of them. Typically, the steps on the adjoint sweep involve more operations and require more storage than the other two. The key difficulty is given by the enormous amount of memory required for the implementation of these methods if all states throughout forward and adjoint sweeps are stored. One of goals of this thesis is to present checkpointing techniques for memory reduced implementation of these methods. For this purpose, the well known aspect of checkpointing has to be extended to a `nested checkpointing` for multiple transversals. The proposed nested reversal schedules drastically reduce the required spatial complexity. The schedules are designed to minimise the overall execution time given a certain total amount of storage for the checkpoints. The proposed scheduling schemes are applied to the memory reduced implementation of the optimal control problem of laser surface hardening and other optimal control problems. / Es wird ein Problem der optimalen Steuerung betrachtet. Die dazugehoerigen Zustandsgleichungen sind mit einer Anfangswertaufgabe definiert. Es existieren zahlreiche numerische Methoden, um Probleme der optimalen Steuerung zu loesen. Der so genannte indirekte Ansatz wird in diesen Thesen detailliert betrachtet. Die indirekten Methoden loesen das aus den Notwendigkeitsbedingungen resultierende Randwertproblem. Das so genannte Pantoja Verfahren beschreibt eine zeiteffiziente schrittweise Berechnung der Newton Richtung fuer diskrete Probleme der optimalen Steuerung. Es gibt mehrere Beziehungen zwischen den unterschiedlichen Mehrzielmethoden und dem Pantoja Verfahren, die in diesen Thesen detailliert zu untersuchen sind. In diesem Zusammenhang wird die aequivalence zwischen dem Pantoja Verfahren und der Mehrzielmethode vom Riccati Typ gezeigt. Ausserdem wird das herkoemlige Pantoja Verfahren dahingehend erweitert, dass die Zustandsgleichungen mit Hilfe einer impliziten numerischen Methode diskretisiert sind. Weiterhin wird das Symplektische Konzept eingefuehrt. In diesem Zusammenhang wird eine geeignete numerische Methode praesentiert, die fuer ein unrestringiertes Problem der optimalen Steuerung angewendet werden kann. In diesen Thesen wird bewiesen, dass diese Methode symplectisch ist. Das iterative Loesen eines Problems der optimalen Steuerung in gewoenlichen Differentialgleichungen mit Hilfe von Pantoja oder Riccati aequivalenten Verfahren fuehrt auf eine Aufeinanderfolge der Durchlaeufetripeln in einem diskretisierten Zeitintervall. Der zweite (adjungierte) Lauf haengt von der Information des ersten (primalen) Laufes, und der dritte (finale) Lauf haeng von den beiden vorherigen ab. Ueblicherweise beinhalten Schritte und Zustaende des adjungierten Laufes wesentlich mehr Operationen und benoetigen auch wesentlich mehr Speicherplatzkapazitaet als Schritte und Zustaende der anderen zwei Durchlaeufe. Das Grundproblem besteht in einer enormen Speicherplatzkapazitaet, die fuer die Implementierung dieser Methoden benutzt wird, falls alle Zustaende des primalen und des adjungierten Durchlaufes zu speichern sind. Ein Ziel dieser Thesen besteht darin, Checkpointing Strategien zu praesentieren, um diese Methoden speichereffizient zu implementieren. Diese geschachtelten Umkehrschemata sind so konstruiert, dass fuer einen gegebenen Speicherplatz die gesamte Laufzeit zur Abarbeitung des Umkehrschemas minimiert wird. Die aufgestellten Umkehrschemata wurden fuer eine speichereffiziente Implementierung von Problemen der optimalen Steuerung angewendet. Insbesondere betrifft dies das Problem einer Oberflaechenabhaertung mit Laserbehandlung.