The advent of cluster computing has resulted in a thrust towards providing software mechanisms for reliability on clusters. The prevalent model for such mechanisms is to take a snapshot of the state of an application, called a checkpoint and commit it to stable storage. This checkpoint has sufficient meta-data, so that if the application fails, it can be restarted from the checkpoint. This operation is called a restore. In order to record a process' complete state, both its volatile and persistent state must be checkpointed. Several libraries exist for checkpointing volatile state. Some of these libraries feature incremental checkpointing, where only the changes since the last checkpoint are recorded in the next checkpoint. Such incremental checkpointing is advantageous since otherwise, the time taken for each successive checkpoint becomes larger and larger. Also, when checkpointing is done in increments, we can restore state to any of the previous checkpoints; a vital feature for adaptive applications. This thesis presents a user-level incremental checkpointing library for files: Metamori. This brings the advantages of incremental memory checkpointing to files as well, thereby providing a low-overhead approach to checkpoint persistent state. Thus, the complete state of an application can now be incrementally checkpointed, as compared to earlier approaches where volatile state was checkpointed incrementally but persistent state had no such facilities. / Master of Science

To make progress in the face of failures, long-running parallel applications need to save their state, known as a checkpoint. Unfortunately, current checkpointing techniques are becoming untenable on large-scale supercomputers. Many applications checkpoint all processes simultaneously--a technique that is easy to implement but often saturates the network and file system, causing a significant increase in checkpoint overhead. This thesis introduces compiler-assisted staggered checkpointing, where processes checkpoint at different places in the application text, thereby reducing contention for the network and file system. This checkpointing technique is algorithmically challenging since the number of possible solutions is enormous and the number of desirable solutions is small, but we have developed a compiler algorithm that both places staggered checkpoints in an application and ensures that the solution is desirable. This algorithm successfully places staggered checkpoints in parallel applications configured to use tens of thousands of processes. For our benchmarks, this algorithm successfully finds and places useful recovery lines that are up to 37% faster for all configurations than recovery lines where all processes write their data at approximately the same time. We also analyze the success of staggered checkpointing by investigating sets of application and system characteristics for which it reduces network and file system contention. We find that for many configurations, staggered checkpointing reduces both checkpointing time and overall execution time. To perform these analyses, we develop an event-driven simulator for large-scale systems that estimates the behavior of the network, global file system, and local hardware using predictive models. Our simulator allows us to accurately study applications that have thousands of processes; it on average predicts execution times as 83% of their measured value. / text

Checkpointing and rollback recovery are well-known techniques for coping with failures in distributed systems. Future generation Supercomputers will be message passing distributed systems consisting of millions of processors. As the number of processors grow, failure rate also grows. Thus, designing efficient checkpointing and recovery algorithms for coping with failures in such large systems is important for these systems to be fully utilized. We presented a novel communication-induced checkpointing algorithm which helps in reducing contention for accessing stable storage to store checkpoints. Under our algorithm, a process involved in a distributed computation can independently initiate consistent global checkpointing by saving its current state, called a tentative checkpoint. Other processes involved in the computation come to know about the consistent global checkpoint initiation through information piggy-backed with the application messages or limited control messages if necessary. When a process comes to know about a new consistent global checkpoint initiation, it takes a tentative checkpoint after processing the message. The tentative checkpoints taken can be flushed to stable storage when there is no contention for accessing stable storage. The tentative checkpoints together with the message logs stored in the stable storage form a consistent global checkpoint. Ad hoc networks consist of a set of nodes that can form a network for communication with each other without the aid of any infrastructure or human intervention. Nodes are energy-constrained and hence routing algorithm designed for these networks should take this into consideration. We proposed two routing protocols for mobile ad hoc networks which prevent nodes from broadcasting route requests unnecessarily during the route discovery phase and hence conserve energy and prevent contention in the network. One is called Triangle Based Routing (TBR) protocol. The other routing protocol we designed is called Routing Protocol with Selective Forwarding (RPSF). Both of the routing protocols greatly reduce the number of control packets which are needed to establish routes between pairs of source nodes and destination nodes. As a result, they reduce the energy consumed for route discovery. Moreover, these protocols reduce congestion and collision of packets due to limited number of nodes retransmitting the route requests.

Yes / Cloud fault tolerance is an important issue in cloud computing platforms and applications. In the event of an unexpected system failure or malfunction, a robust fault-tolerant design may allow the cloud to continue functioning correctly possibly at a reduced level instead of failing completely. To ensure high availability of critical cloud services, the application execution and hardware performance, various fault tolerant techniques exist for building self-autonomous cloud systems. In comparison to current approaches, this paper proposes a more robust and reliable architecture using optimal checkpointing strategy to ensure high system availability and reduced system task service finish time. Using pass rates and virtualised mechanisms, the proposed Smart Failover Strategy (SFS) scheme uses components such as Cloud fault manager, Cloud controller, Cloud load balancer and a selection mechanism, providing fault tolerance via redundancy, optimized selection and checkpointing. In our approach, the Cloud fault manager repairs faults generated before the task time deadline is reached, blocking unrecoverable faulty nodes as well as their virtual nodes. This scheme is also able to remove temporary software faults from recoverable faulty nodes, thereby making them available for future request. We argue that the proposed SFS algorithm makes the system highly fault tolerant by considering forward and backward recovery using diverse software tools. Compared to existing approaches, preliminary experiment of the SFS algorithm indicate an increase in pass rates and a consequent decrease in failure rates, showing an overall good performance in task allocations. We present these results using experimental validation tools with comparison to other techniques, laying a foundation for a fully fault tolerant IaaS Cloud environment.

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

With the advent of Chip Multi Processors (CMPs), improving performance relies on the programmers/compilers to expose thread level parallelism to the underlying hardware. Unfortunately, this is a difficult and error-prone process for the programmers, while state of the art compiler techniques are unable to provide significant benefits for many classes of applications. An interesting alternative is offered by systems that support Thread Level Speculation (TLS), which relieve the programmer and compiler from checking for thread dependencies and instead use the hardware to enforce them. Unfortunately, data misspeculation results in a high cost since all the intermediate results have to be discarded and threads have to roll back to the beginning of the speculative task. For this reason intermediate checkpointing of the state of the TLS threads has been proposed. When the violation does occur, we now have to roll back to a checkpoint before the violating instruction and not to the start of the task. However, previous work omits study of the microarchitectural details and implementation issues that are essential for effective checkpointing. Further, checkpoints have only been proposed and evaluated for a narrow class of benchmarks. This thesis studies checkpoints on a state of the art TLS system running a variety of benchmarks. The mechanisms required for checkpointing and the costs associated are described. Hardware modifications required for making checkpointed execution efficient in time and power are proposed and evaluated. Further, the need for accurately identifying suitable points for placing checkpoints is established. Various techniques for identifying these points are analysed in terms of both effectiveness and viability. This includes an extensive evaluation of data dependence prediction techniques. The results show that checkpointing thread level speculative execution results in consistent power savings, and for many benchmarks leads to speedups as well.

Distributed Transactional Memory (DTM) is an emerging synchronization abstraction thatpromises to alleviate the scalability, programmability, and composability challenges of lock-based distributed synchronization. With DTM, programmers organize code that read/writeshared memory objects, both local and remote, as memory transactions. An underlying DTMframework guarantees atomicity, isolation, and consistency properties for those transactionsthrough speculative concurrency control. In DTM, restarting an aborted transaction from the beginning can degrade performance astransactional conflicts may have occurred in the later part of the transaction, wasting work.The partial abort method alleviates this difficulty by enabling a transaction to be rolled backto the point where objects in the transaction's read-set and write-set are still consistent.In this thesis, we present protocols for supporting partial aborts in QR-DTM, a fault-tolerant DTM that uses quorum protocols for distributed concurrency control in the presence offailures. We focus on two partial abort models: closed nesting, which allows transactions to be nested and inner transactions to be rolled back without rolling back outer transactions;and checkpointing, which allows the transaction state to be saved in checkpoints throughout execution and transactions to be rolled back to those checkpoints. We present protocols that support these partial abort models in QR-DTM, called QR-CN and QR-CHK. We implemented these protocols and conducted experimental studies using macro-benchmarks(e.g., distributed versions of STAMP benchmark), and micro-benchmarks (e.g., distributed data structures). Our studies reveal that QR-CN improves throughput by as much as 101% over flat nesting in specific cases, with an average improvement of 53%. We also develop QR-ACN, a framework that automatically decomposes programmer-writtentransactions into closed nested transactions, at run-time, to improve performance. The composition of a closed nested transaction depends on the contention of objects, which can change at run-time depending upon the workload at hand. Our implementation and experimental studies reveal that QR-ACN consistently outperforms flat nesting by an averageof 51% on benchmarks including TPC-C. False conflicts occur when high-level operations, even though semantically independent, traverse the same set of objects during transaction execution. Such conflicts can lead torepeated aborts, increasing transaction execution time and degrading performance, which can be significant in DTM, since transaction execution also includes network communication. ii We consider the approach of reducing validation cost for resolving false conflicts. We presentthree protocols for reducing validation cost in DTM. Our first protocol, QR-ON, incorporatesthe open nesting model into QR-DTM. Open nesting allows inner-nested transactions tocommit independently of their parent transaction, releasing objects in the transaction read-set and write-set early, minimizing aborts due to false conflicts. We then present QR-OON,in which open-nested transactions commit asynchronously so that subsequent transactionscan proceed without waiting for the commit of previous transactions. Finally, we presentan early release methodology, QR-ER, in which objects that do not affect the final state ofthe shared data are dropped from the transaction's read-set, which avoids false conflicts andreduces communication costs. Our implementation and experimental studies revealed that QR-OON outperforms QR-ON by up to 43%, and that QR-ER outperforms QR-ON and QR-OON by up to 10% on micro- and macro-benchmarks. / Master of Science

Bei der Berechnung von Adjungierten, zum Debuggen und für ähnliche Anwendungen kann man die Umkehr der entsprechenden Programmauswertung verwenden. Der einfachste Ansatz, nämlich das Mitschreiben einer kompletten Mitschrift der Vorwärtsrechnung, welche anschließend rückwärts gelesen wird, verursacht einen enormen Speicherplatzbedarf. Als Alternative dazu kann man die Mitschrift auch stückweise erzeugen, indem die Programmauswertung von passend gewählten Checkpoints wiederholt gestartet wird. Das Ziel der Arbeit ist die Minimierung des von der Programmumkehr verursachten Zeit- und Speicherplatzbedarfs. Dieser wird gemessen in Auswertungswiederholungen bzw. verwendeten Checkpoints. Optimale Umkehrschemata werden für Ein- und Mehr-Schritt-Verfahren entwickelt, welche zum Beispiel bei der Diskretisierung einer gewöhnlichen Differentialgleichung Verwendung finden. Desweiteren erfolgte die Entwicklung von parallelen Umkehrschemata, d. h. mehrere Prozessoren werden für die Umkehrung der Programmauswertung eingesetzt. Diese zusätzlichen Prozessoren dienen dazu, die wiederholten Berechnungen des Programms zu parallelisieren, so daß ein Prozessor die Rückwartsrechnung ohne Unterbrechung durchführen kann. Sowohl für die seriellen als auch für die parallelen Umkehrschemata wurde gezeigt, daß die Länge der umzukehrenden Programmauswertung exponentiell in Abhängigkeit von der Zahl der verwendeten Checkpoints und der Zahl der wiederholten Auswertungen bzw. verwendeten Prozessoren wächst. / For adjoint calculations, parameter estimation, and similar purposes one may need to reverse the execution of a computer program. The simplest option is to record a complete execution log and then to read it backwards. This requires massive amounts of storage. Instead one may generate the execution log piecewise by restarting the ``forward'' calculation repeatedly from suitably placed checkpoints. The basic structure of the resulting reversal schedules is illustrated. Various strategies are analysed with respect to the resulting temporal and spatial complexity on serial and parallel machines. For serial machines known optimal compromises between operations count and memory requirement are explained, and they are extended to more general situations. For program execution reversal on multi-processors the new challenges and demands on an optimal reversal schedule are described. We present parallel reversal schedules that are provably optimal with regards to the number of concurrent processes and the total amount of memory required.

OpenMP et MPI sont devenus les outils standards pour développer des programmes parallèles sur une architecture à mémoire partagée et à mémoire distribuée respectivement. Comparé à MPI, OpenMP est plus facile à utiliser. Ceci est dû au fait qu’OpenMP génère automatiquement le code parallèle et synchronise les résultats à l’aide de directives, clauses et fonctions d’exécution, tandis que MPI exige que les programmeurs fassent ce travail manuellement. Par conséquent, des efforts ont été faits pour porter OpenMP sur les architectures à mémoire distribuée. Cependant, à l’exclusion de CAPE, aucune solution ne satisfait les deux exigences suivantes: 1) être totalement conforme à la norme OpenMP et 2) être hautement performant. CAPE (Checkpointing-Aided Parallel Execution) est un framework qui traduit et fournit automatiquement des fonctions d’exécution pour exécuter un programme OpenMP sur une architecture à mémoire distribuée basé sur des techniques de checkpoint. Afin d’exécuter un programme OpenMP sur un système à mémoire distribuée, CAPE utilise un ensemble de modèles pour traduire le code source OpenMP en code source CAPE, puis le code source CAPE est compilé par un compilateur C/C++ classique. Fondamentalement, l’idée de CAPE est que le programme s’exécute d’abord sur un ensemble de nœuds du système, chaque nœud fonctionnant comme un processus. Chaque fois que le programme rencontre une section parallèle, le maître distribue les tâches aux processus esclaves en utilisant des checkpoints incrémentaux discontinus (DICKPT). Après l’envoi des checkpoints, le maître attend les résultats renvoyés par les esclaves. L’étape suivante au niveau du maître consiste à recevoir et à fusionner le résultat des checkpoints avant de les injecter dans sa mémoire. Les nœuds esclaves quant à eux reçoivent les différents checkpoints, puis l’injectent dans leur mémoire pour effectuer le travail assigné. Le résultat est ensuite renvoyé au master en utilisant DICKPT. À la fin de la région parallèle, le maître envoie le résultat du checkpoint à chaque esclave pour synchroniser l’espace mémoire du programme. Dans certaines expériences, CAPE a montré des performances élevées sur les systèmes à mémoire distribuée et constitue une solution viable entièrement compatible avec OpenMP. Cependant, CAPE reste en phase de développement, ses checkpoints et son modèle d’exécution devant être optimisés pour améliorer les performances, les capacités et la fiabilité. Cette thèse vise à présenter les approches proposées pour optimiser et améliorer la capacité des checkpoints, concevoir et mettre en œuvre un nouveau modèle d’exécution, et améliorer la capacité de CAPE. Tout d’abord, nous avons proposé une arithmétique sur les checkpoints qui modélise la structure de leurs données et ses opérations. Cette modélisation contribue à optimiser leur taille et à réduire le temps nécessaire à la fusion, tout en améliorant leur capacité. Deuxièmement, nous avons développé TICKPT (Time-Stamp Incremental Checkpointing) une implémentation de l’arithmétique sur les checkpoints. TICKPT est une amélioration de DICKPT, il a ajouté l’horodatage aux checkpoints pour en identifier l’ordre. L’analyse et les expériences comparées montrent TICKPT sont non seulement plus petites, mais qu’ils ont également moins d’impact sur les performances du programme. Troisièmement, nous avons conçu et implémenté un nouveau modèle d’exécution et de nouveaux prototypes pour CAPE basés sur TICKPT. Le nouveau modèle d’exécution permet à CAPE d’utiliser les ressources efficacement, d’éviter les risques de goulots d’étranglement et de satisfaire à l’exigence des les conditions de Bernstein. Au final, ces approches améliorent significativement les performances de CAPE, ses capacités et sa fiabilité. Le partage des données implémenté sur CAPE et basé sur l’arithmétique sur des checkpoints est ouvert et basé sur TICKPT. Cela démontre également la bonne direction que nous avons prise et rend CAPE plus complet / OpenMP and MPI have become the standard tools to develop parallel programs on shared-memory and distributed-memory architectures respectively. As compared to MPI, OpenMP is easier to use. This is due to the ability of OpenMP to automatically execute code in parallel and synchronize results using its directives, clauses, and runtime functions while MPI requires programmers do all this manually. Therefore, some efforts have been made to port OpenMP on distributed-memory architectures. However, excluding CAPE, no solution has successfully met both requirements: 1) to be fully compliant with the OpenMP standard and 2) high performance. CAPE stands for Checkpointing-Aided Parallel Execution. It is a framework that automatically translates and provides runtime functions to execute OpenMP program on distributed-memory architectures based on checkpointing techniques. In order to execute an OpenMP program on distributed-memory system, CAPE uses a set of templates to translate OpenMP source code to CAPE source code, and then, the CAPE source code is compiled by a C/C++ compiler. This code can be executed on distributed-memory systems under the support of the CAPE framework. Basically, the idea of CAPE is the following: the program first run on a set of nodes on the system, each node being executed as a process. Whenever the program meets a parallel section, the master distributes the jobs to the slave processes by using a Discontinuous Incremental Checkpoint (DICKPT). After sending the checkpoints, the master waits for the returned results from the slaves. The next step on the master is the reception and merging of the resulting checkpoints before injecting them into the memory. For slave nodes, they receive different checkpoints, and then, they inject it into their memory to compute the divided job. The result is sent back to the master using DICKPTs. At the end of the parallel region, the master sends the result of the checkpoint to every slaves to synchronize the memory space of the program as a whole. In some experiments, CAPE has shown very high-performance on distributed-memory systems and is a viable and fully compatible with OpenMP solution. However, CAPE is in the development stage. Its checkpoint mechanism and execution model need to be optimized in order to improve the performance, ability, and reliability. This thesis aims at presenting the approaches that were proposed to optimize and improve checkpoints, design and implement a new execution model, and improve the ability for CAPE. First, we proposed arithmetics on checkpoints, which aims at modeling checkpoint’s data structure and its operations. This modeling contributes to optimize checkpoint size and reduces the time when merging, as well as improve checkpoints capability. Second, we developed TICKPT which stands for Time-stamp Incremental Checkpointing as an instance of arithmetics on checkpoints. TICKPT is an improvement of DICKPT. It adds a timestamp to checkpoints to identify the checkpoints order. The analysis and experiments to compare it to DICKPT show that TICKPT do not only provide smaller in checkpoint size, but also has less impact on the performance of the program using checkpointing. Third, we designed and implemented a new execution model and new prototypes for CAPE based on TICKPT. The new execution model allows CAPE to use resources efficiently, avoid the risk of bottlenecks, overcome the requirement of matching the Bernstein’s conditions. As a result, these approaches make CAPE improving the performance, ability as well as reliability. Four, Open Data-sharing attributes are implemented on CAPE based on arithmetics on checkpoints and TICKPT. This also demonstrates the right direction that we took, and makes CAPE more complete

Grades computacionais oportunistas utilizam recursos ociosos de máquinas compartilhadas para executar aplicações que necessitam de um alto poder computacional e/ou trabalham com grandes quantidades de dados. Mas a execução de aplicações paralelas computacionalmente intensivas em ambientes dinâmicos e heterogêneos, como grades computacionais oportunistas, é uma tarefa difícil. Máquinas podem falhar, ficar inacessíveis ou passar de ociosas para ocupadas inesperadamente, comprometendo a execução de aplicações. Um mecanismo de tolerância a falhas que dê suporte a arquiteturas heterogêneas é um importante requisito para estes sistemas. Neste trabalho, analisamos, implementamos e avaliamos um mecanismo de tolerância a falhas baseado em checkpointing para aplicações paralelas em grades computacionais oportunistas. Este mecanismo permite o monitoramento de execuções e a migração de aplicações entre nós heterogêneos da grade. Mas além da execução, é preciso gerenciar e armazenar os dados gerados e utilizados por estas aplicações. Desejamos uma infra-estrutura de armazenamento de dados de baixo custo e que utilize o espaço livre em disco de máquinas compartilhadas da grade. Devemos utilizar somente os ciclos ociosos destas máquinas para armazenar e recuperar dados, de modo que um sistema de armazenamento distribuído que as utilize deve ser redundante e tolerante a falhas. Para resolver o problema do armazenamento de dados em grades oportunistas, projetamos, implementamos e avaliamos o middleware OppStore. Este middleware provê armazenamento distribuído e confiável de dados, que podem ser acessados de qualquer máquina da grade. As máquinas são organizadas em aglomerados, que são conectados por uma rede peer-to-peer auto-organizável e tolerante a falhas. Dados são codificados em fragmentos redundantes antes de serem armazenados, de modo que arquivos podem ser reconstruídos utilizando apenas um subconjunto destes fragmentos. Finalmente, para lidar com a heterogeneidade dos recursos, desenvolvemos uma extensão ao protocolo de roteamento em redes peer-to-peer Pastry. Esta extensão adiciona balanceamento de carga e suporte à heterogeneidade de máquinas ao protocolo Pastry. / Opportunistic computational grids use idle resources from shared machines to execute applications that need large amounts of computational power and/or deal with large amounts of data. But executing computationally intensive parallel applications in dynamic and heterogeneous environments, such as opportunistic grids, is a daunting task. Machines may fail, become inaccessible, or change from idle to occupied unexpectedly, compromising the application execution. A fault tolerance mechanism that supports heterogeneous architectures is an important requisite for such systems. In this work, we analyze, implement and evaluate a checkpointing-based fault tolerance mechanism for parallel applications running on opportunistic grids. The mechanism monitors application execution and allows the migration of applications between heterogeneous nodes of the grid. But besides application execution, it is necessary to manage data generated and used by those applications. We want a low cost data storage infrastructure that utilizes the unused disk space of grid shared machines. The system should use the machines to store and recover data only during their idle periods, requiring the system to be redundant and fault-tolerant. To solve the data storage problem in opportunistic grids, we designed, implemented and evaluated the OppStore middleware. This middleware provides reliable distributed storage for application data, which can be accessed from any machine in the grid. The machines are organized in clusters, connected by a self-organizing and fault-tolerant peer-to-peer network. During storage, data is codified into redundant fragments, allowing the reconstruction of the original file using only a subset of those fragments. Finally, to deal with resource heterogeneity, we developed an extension to the Pastry peer-to-peer routing substrate, enabling heterogeneity-aware load-balancing message routing.