The colossal amounts of data generated daily are increasing exponentially at a never-before-seen pace. A variety of applications—including stock trading, banking systems, health-care, Internet of Things (IoT), and social media networks, among others—have created an unprecedented volume of real-time stream data estimated to reach billions of terabytes in the near future. As a result, we are currently living in the so-called Big Data era and witnessing a transition to the so-called IoT era. Enterprises and organizations are tackling the challenge of interpreting the enormous amount of raw data streams to achieve an improved understanding of data, and thus make efficient and well-informed decisions (i.e., data-driven decisions). Researchers have designed distributed data stream processing systems that can directly process data in near real-time. To extract valuable information from raw data streams, analysts need to create and implement data stream processing applications structured as a directed acyclic graphs (DAG). The infrastructure of distributed data stream processing systems, as well as the various requirements of stream applications, impose new challenges. Cluster heterogeneity in a distributed environment results in different cluster resources for task execution and data transmission, which make the optimal scheduling algorithms an NP-complete problem. Scheduling streaming applications plays a key role in optimizing system performance, particularly in maximizing the frame-rate, or how many instances of data sets can be processed per unit of time. The scheduling algorithm must consider data locality, resource heterogeneity, and communicational and computational latencies. The latencies associated with the bottleneck from computation or transmission need to be minimized when mapped to the heterogeneous and distributed cluster resources. Recent work on task scheduling for distributed data stream processing systems has a number of limitations. Most of the current schedulers are not designed to manage heterogeneous clusters. They also lack the ability to consider both task and machine characteristics in scheduling decisions. Furthermore, current default schedulers do not allow the user to control data locality aspects in application deployment.In this thesis, we investigate the problem of scheduling streaming applications on a heterogeneous cluster environment and develop the maximum throughput scheduler algorithm (MT-Scheduler) for streaming applications. The proposed algorithm uses a dynamic programming technique to efficiently map the application topology onto a heterogeneous distributed system based on computing and data transfer requirements, while also taking into account the capacity of underlying cluster resources. The proposed approach maximizes the system throughput by identifying and minimizing the time incurred at the computing/transfer bottleneck. The MT-Scheduler supports scheduling applications that are structured as a DAG, such as Amazon Timestream, Google Millwheel, and Twitter Heron. We conducted experiments using three Storm microbenchmark topologies in both simulated and real Apache Storm environments. To evaluate performance, we compared the proposed MT-Scheduler with the simulated round-robin and the default Storm scheduler algorithms. The results indicated that the MT-Scheduler outperforms the default round-robin approach in terms of both average system latency and throughput.

Multi- and many-core processors are becoming increasingly popular in embedded systems. Many of these processors now feature hardware virtualization capabilities, such as the ARM Cortex A15, and x86 processors with Intel VT-x or AMD-V support. Hardware virtualization offers opportunities to partition physical resources, including processor cores, memory and I/O devices amongst guest virtual machines. Mixed criticality systems and services can then co-exist on the same platform in separate virtual machines. However, traditional virtual machine systems are too expensive because of the costs of trapping into hypervisors to multiplex and manage machine physical resources on behalf of separate guests. For example, hypervisors are needed to schedule separate VMs on physical processor cores. Additionally, traditional hypervisors have memory footprints that are often too large for many embedded computing systems. This dissertation presents the design of the Quest-V separation kernel, which partitions services of different criticality levels across separate virtual machines, or sandboxes. Each sandbox encapsulates a subset of machine physical resources that it manages without requiring intervention of a hypervisor. In Quest-V, a hypervisor is not needed for normal operation, except to bootstrap the system and establish communication channels between sandboxes. This approach not only reduces the memory footprint of the most privileged protection domain, it removes it from the control path during normal system operation, thereby heightening security.

La dernière décennie a vu l’émergence de systèmes parallèles pour l’analyse de grosse quantités de données (DISC) , tels que Hadoop, et la demande qui en résulte pour les politiques de gestion des ressources, pouvant fournir des temps de réponse rapides ainsi qu’équité. Actuellement, les schedulers pour les systèmes de DISC sont axées sur l’équité, sans optimiser les temps de réponse. Les meilleures pratiques pour surmonter ce problème comprennent une intervention manuelle et une politique de planification ad-hoc , qui est sujette aux erreurs et qui est difficile à adapter aux changements. Dans cette thèse, nous nous concentrons sur la planification basée sur la taille pour les systèmes DISC. La principale contribution de ce travail est le scheduler dit Hadoop Fair Sojourn Protocol (HFSP), un ordonnanceur préemptif basé sur la taille qui tient en considération le vieillissement, ayant comme objectifs de fournir l’équité et des temps de réponse réduits. Hélas, dans les systèmes DISC, les tailles des job d’analyse de données ne sont pas connus a priori, donc, HFSP comprends un module d’estimation de taille, qui calcule une approximation et qui affine cette estimation au fur et a mesure du progrès d’un job. Nous démontrons que l’impact des erreurs d’estimation sur les politiques fondées sur la taille n’est pas significatif. Pour cette raison, et en vertu d’être conçu autour de l’idée de travailler avec des tailles estimées, HFSP est tolérant aux erreurs d’estimation de la taille des jobs. Nos résultats expérimentaux démontrent que, dans un véritable déploiement Hadoop avec des charges de travail réalistes, HFSP est plus performant que les politiques de scheduling existantes, a la fois en terme de temps de réponse et d’équité. En outre, HFSP maintiens ses bonnes performances même lorsque le cluster de calcul est lourdement chargé, car il focalises les ressources sur des jobs ayant priorité. HFSP est une politique préventive: la préemption dans un système DISC peut être mis en œuvre avec des techniques différentes. Les approches actuellement disponibles dans Hadoop ont des lacunes qui ont une incidence sur les performances du système. Par conséquence, nous avons mis en œuvre une nouvelle technique de préemption, appelé suspension, qui exploite le système d’exploitation pour effectuer la préemption d’une manière qui garantie une faible latence sans pénaliser l’avancement des jobs a faible priorité. / The past decade have seen the rise of data-intensive scalable computing (DISC) systems, such as Hadoop, and the consequent demand for scheduling policies to manage their resources, so that they can provide quick response times as well as fairness. Schedulers for DISC systems are usually focused on the fairness, without optimizing the response times. The best practices to overcome this problem include a manual and ad-hoc control of the scheduling policy, which is error-prone and difficult to adapt to changes. In this thesis we focus on size-based scheduling for DISC systems. The main contribution of this work is the Hadoop Fair Sojourn Protocol (HFSP) scheduler, a size-based preemptive scheduler with aging; it provides fairness and achieves reduced response times thanks to its size-based nature. In DISC systems, job sizes are not known a-priori: therefore, HFSP includes a job size estimation module, which computes approximated job sizes and refines these estimations as jobs progress. We show that the impact of estimation errors on the size-based policies is not signifi- cant, under conditions which are verified in a system such as Hadoop. Because of this, and by virtue of being designed around the idea of working with estimated sizes, HFSP is largely tolerant to job size estimation errors. Our experimental results show that, in a real Hadoop deployment and with realistic workloads, HFSP performs better than the built-in scheduling policies, achieving both fairness and small mean response time. Moreover, HFSP maintains its good performance even when the cluster is heavily loaded, by focusing the resources to few selected jobs with the smallest size. HFSP is a preemptive policy: preemption in a DISC system can be implemented with different techniques. Approaches currently available in Hadoop have shortcomings that impact on the system performance. Therefore, we have implemented a new preemption technique, called suspension, that exploits the operating system primitives to implement preemption in a way that guarantees low latency without penalizing low-priority jobs.

Nous vivons dans un monde de big data, où les données sont générées en grand volume, grande vitesse et grande variété. Le big data apportent des valeurs et des avantages énormes, de sorte que l’analyse des données est devenue un facteur essentiel de succès commercial dans tous les secteurs. Cependant, si les données ne sont pas analysées assez rapidement, les bénéfices de big data seront limités ou même perdus. Malgré l’existence de nombreux systèmes modernes d’analyse de données à grande échelle, la préparation des données est le processus le plus long de l’analyse des données, n’a pas encore reçu suffisamment d’attention. Dans cette thèse, nous étudions le problème de la façon d’accélérer la préparation des données pour le big data d’analyse. En particulier, nous nous concentrons sur deux grandes étapes de préparation des données, le chargement des données et le nettoyage des données. Comme première contribution de cette thèse, nous concevons DiNoDB, un système SQL-on-Hadoop qui réalise l’exécution de requêtes à vitesse interactive sans nécessiter de chargement de données. Les applications modernes impliquent de lourds travaux de traitement par lots sur un grand volume de données et nécessitent en même temps des analyses interactives ad hoc efficaces sur les données temporaires générées dans les travaux de traitement par lots. Les solutions existantes ignorent largement la synergie entre ces deux aspects, nécessitant de charger l’ensemble des données temporaires pour obtenir des requêtes interactives. En revanche, DiNoDB évite la phase coûteuse de chargement et de transformation des données. L’innovation importante de DiNoDB est d’intégrer à la phase de traitement par lots la création de métadonnées que DiNoDB exploite pour accélérer les requêtes interactives. La deuxième contribution est un système de flux distribué de nettoyage de données, appelé Bleach. Les approches de nettoyage de données évolutives existantes s’appuient sur le traitement par lots pour améliorer la qualité des données, qui demandent beaucoup de temps. Nous ciblons le nettoyage des données de flux dans lequel les données sont nettoyées progressivement en temps réel. Bleach est le premier système de nettoyage qualitatif de données de flux, qui réalise à la fois la détection des violations en temps réel et la réparation des données sur un flux de données sale. Il s’appuie sur des structures de données efficaces, compactes et distribuées pour maintenir l’état nécessaire pour nettoyer les données et prend également en charge la dynamique des règles. Nous démontrons que les deux systèmes résultants, DiNoDB et Bleach, ont tous deux une excellente performance par rapport aux approches les plus avancées dans nos évaluations expérimentales, et peuvent aider les chercheurs à réduire considérablement leur temps consacré à la préparation des données. / We are living in a big data world, where data is being generated in high volume, high velocity and high variety. Big data brings enormous values and benefits, so that data analytics has become a critically important driver of business success across all sectors. However, if the data is not analyzed fast enough, the benefits of big data will be limited or even lost. Despite the existence of many modern large-scale data analysis systems, data preparation which is the most time-consuming process in data analytics has not received sufficient attention yet. In this thesis, we study the problem of how to accelerate data preparation for big data analytics. In particular, we focus on two major data preparation steps, data loading and data cleaning. As the first contribution of this thesis, we design DiNoDB, a SQL-on-Hadoop system which achieves interactive-speed query execution without requiring data loading. Modern applications involve heavy batch processing jobs over large volume of data and at the same time require efficient ad-hoc interactive analytics on temporary data generated in batch processing jobs. Existing solutions largely ignore the synergy between these two aspects, requiring to load the entire temporary dataset to achieve interactive queries. In contrast, DiNoDB avoids the expensive data loading and transformation phase. The key innovation of DiNoDB is to piggyback on the batch processing phase the creation of metadata, that DiNoDB exploits to expedite the interactive queries. The second contribution is a distributed stream data cleaning system, called Bleach. Existing scalable data cleaning approaches rely on batch processing to improve data quality, which are very time-consuming in nature. We target at stream data cleaning in which data is cleaned incrementally in real-time. Bleach is the first qualitative stream data cleaning system, which achieves both real-time violation detection and data repair on a dirty data stream. It relies on efficient, compact and distributed data structures to maintain the necessary state to clean data, and also supports rule dynamics. We demonstrate that the two resulting systems, DiNoDB and Bleach, both of which achieve excellent performance compared to state-of-the-art approaches in our experimental evaluations, and can help data scientists significantly reduce their time spent on data preparation.

付記する学位プログラム名: グローバル生存学大学院連携プログラム / 京都大学 / 0048 / 新制・課程博士 / 博士(工学) / 甲第22771号 / 工博第4770号 / 新制||工||1746(附属図書館) / 京都大学大学院工学研究科機械理工学専攻 / (主査)教授 松野 文俊, 教授 椹木 哲夫, 教授 泉田 啓 / 学位規則第4条第1項該当 / Doctor of Philosophy (Engineering) / Kyoto University / DFAM

Large-scale distributed computing systems such as data centers are hosted on heterogeneous and networked servers that execute in a dynamic and uncertain operating environment, caused by factors such as time-varying user workload and various failures. Therefore, achieving stringent quality-of-service goals is a challenging task, requiring a comprehensive approach to performance control, fault diagnosis, and failure recovery. This work presents a model-based approach for fault management, which integrates limited lookahead control (LLC), diagnosis, and fault-tolerance concepts that: (1) enables systems to adapt to environment variations, (2) maintains the availability and reliability of the system, (3) facilitates system recovery from failures. We focused on memory leak errors in this thesis. A characterization function is designed to detect memory leaks. Then, a LLC is applied to enable the computing system to adapt efficiently to variations in the workload, and to enable the system recover from memory leaks and maintain functionality.

No description available.

No description available.

Big data are datasets that is very large and computational complex. With an increasing volume of data the time a trivial processing task can be challenging. Companies collects data at a fast rate but knowing what to do with the data can be hard. A search engine is a system that indexes data making it efficiently queryable by users. When a bug occurs in a computer system log data is consulted in order to understand why, but processing big log data can take a long time. The purpose of this thesis is to investigate, compare and implement a distributed search engine that can prepare log data for analysis, which will make it easier for a developer to investigate bugs. There are three popular search engines: Apache Lucene, Elasticsearch and Apache Solr. Elasticsearch and Apache Solr are built as distributed systems making them capable of handling big data. Requirements was established through interviews. Big log data of totally 40 GB was provided that would be indexed in the selected search engine. The log data provided was generated in a proprietary binary format and it had to be decoded before. The distributed search engines was evaluated based on: Distributed architecture, text analysis, indexing and querying. Elasticsearch was selected for implementation. A cluster was set up on Amazon Web Services and tests was executed in order to determine how different configurations performed. An indexing software was written that would transfer data to the cluster. Results was verified through a case-study with participants of the stakeholder. / Stordata är en datamängd som är mycket stora och komplexa att göra beräkningar på. När en datamängd ökar blir en trivial bearbetningsuppgift betydligt mera utmanande. Företagen samlar idag in data i allt snabbare takt men det är svårt att veta exakt vad man ska göra med den data. En sökmotor är ett system som indexerar data och gör det effektivt att för användare att söka i det. När ett fel inträffar i ett datorsystem går utvecklare igenom loggdata för att få en insikt i varför, men det kan ta lång tid att söka igenom en stor mängd loggdata. Syftet med denna avhandling är att undersöka, jämföra och implementera en distribuerad sökmotor som kan förbereda loggdata för analys, vilket gör det lättare för utvecklare att undersöka buggar. Det finns tre populära sökmotorer: Apache Lucene, Elasticsearch och Apache Solr. Elasticsearch och Apache Solr är byggda som distribuerade system och kan därav hantera stordata. Krav fastställdes genom intervjuer. En stor mängd loggdata på totalt 40 GB indexerades i den valda sökmotorn. Den loggdata som användes genererades i en proprietär binärt format som behövdes avkodas för att kunna användas. De distribuerade sökmotorerna utvärderades utifrån kriterierna: Distribuerad arkitektur, textanalys, indexering och förfrågningar. Elasticsearch valdes för att implementeras. Ett kluster sattes upp på Amazon Web Services och test utfördes för att bestämma hur olika konfigurationer presterade. En indexeringsprogramvara skrevs som skulle överföra data till klustret. Resultatet verifierades genom en studie med deltagare från intressenten.

While existing grid environments cater to specific needs of a particular user community, we need to go beyond them and consider general-purpose large-scale distributed systems consisting of large collections of heterogeneous computers and communication systems shared by a large user population with very diverse requirements. Coordination, matchmaking, and resource allocation are among the essential functions of large-scale distributed systems. Although deterministic approaches for coordination, matchmaking, and resource allocation have been well studied, they are not suitable for large-scale distributed systems due to the large-scale, the autonomy, and the dynamics of the systems. We have to seek for nondeterministic solutions for large-scale distributed systems. In this dissertation we describe our work on a coordination service, a matchmaking service, and a macro-economic resource allocation model for large-scale distributed systems. The coordination service coordinates the execution of complex tasks in a dynamic environment, the matchmaking service supports finding the appropriate resources for users, and the macro-economic resource allocation model allows a broker to mediate resource providers who want to maximize their revenues and resource consumers who want to get the best resources at the lowest possible price, with some global objectives, e.g., to maximize the resource utilization of the system.