Data streams in the form of potentially unbounded sequences of tuples arise naturally in a large variety of domains including finance markets, sensor networks, social media, and network traffic management. The increasing number of applications that require processing data streams with high throughput and low latency have promoted the development of data stream processing systems (DSPS). A DSPS processes data streams with continuous queries, which are issued once and return query results to users continuously as new tuples arrive. For stream-based applications, both the query-execution performance (in terms of, e.g., throughput and end-to-end latency) and the quality of produced query results (in terms of, e.g., accuracy and completeness) are important. However, a DSPS often needs to make tradeoffs between these two requirements, either because of the data imperfection within the streams, or because of the limited computation capacity of the DSPS itself. Performance versus result-quality tradeoffs caused by data imperfection are inevitable, because the quality of the incoming data is beyond the control of a DSPS, whereas tradeoffs caused by system limitations can be alleviated—even erased—by enhancing the DSPS itself. This dissertation seeks to advance the state of the art on handling the performance versus result-quality tradeoffs in data stream processing caused by the above two aspects of reasons. For tradeoffs caused by data imperfection, this dissertation focuses on the typical data-imperfection problem of stream disorder and proposes the concept of quality-driven disorder handling (QDDH). QDDH enables a DSPS to make flexible and user-configurable tradeoffs between the end-to-end latency and the query-result quality when dealing with stream disorder. Moreover, compared to existing disorder handling approaches, QDDH can significantly reduce the end-to-end latency, and at the same time provide users with desired query-result quality. In this dissertation, a generic buffer-based QDDH framework and three instantiations of the generic framework for distinct query types are presented. For tradeoffs caused by system limitations, this dissertation proposes a system-enhancement approach that combines the row-oriented and the column-oriented data layout and processing techniques in data stream processing to improve the throughput. To fully exploit the potential of such hybrid execution of continuous queries, a static, cost-based query optimizer is introduced. The optimizer works at the operator level and takes the unique property of execution plans of continuous queries—feasibility—into account.

Virtualization is a mature technology in server and desktop environments where multiple systems are consolidate onto a single physical hardware platform, increasing the utilization of todays multi-core systems as well as saving resources such as energy, space and costs compared to multiple single systems. Looking at embedded environments reveals that many systems use multiple separate computing systems inside, including requirements for real-time and isolation properties. For example, modern high-comfort cars use up to a hundred embedded computing systems. Consolidating such diverse configurations promises to save resources such as energy and weight. In my work I propose a secure software architecture that allows consolidating multiple embedded software systems with timing constraints. The base of the architecture builds a microkernel-based operating system that supports a variety of different virtualization approaches through a generic interface, supporting hardware-assisted virtualization and paravirtualization as well as multiple architectures. Studying guest systems with latency constraints with regards to virtualization showed that standard techniques such as high-frequency time-slicing are not a viable approach. Generally, guest systems are a combination of best-effort and real-time work and thus form a mixed-criticality system. Further analysis showed that such systems need to export relevant internal scheduling information to the hypervisor to support multiple guests with latency constraints. I propose a mechanism to export those relevant events that is secure, flexible, has good performance and is easy to use. The thesis concludes with an evaluation covering the virtualization approach on the ARM and x86 architectures and two guest operating systems, Linux and FreeRTOS, as well as evaluating the export mechanism.

Event stream processing (ESP) applications enable the real-time processing of continuous flows of data. Algorithmic trading, network monitoring, and processing data from sensor networks are good examples of applications that traditionally rely upon ESP systems. In addition, technological advances are resulting in an increasing number of devices that are network enabled, producing information that can be automatically collected and processed. This increasing availability of on-line data motivates the development of new and more sophisticated applications that require low-latency processing of large volumes of data. ESP applications are composed of an acyclic graph of operators that is traversed by the data. Inside each operator, the events can be transformed, aggregated, enriched, or filtered out. Some of these operations depend only on the current input events, such operations are called stateless. Other operations, however, depend not only on the current event, but also on a state built during the processing of previous events. Such operations are, therefore, named stateful. As the number of ESP applications grows, there are increasingly strong requirements, which are often difficult to satisfy. In this dissertation, we address two challenges created by the use of stateful operations in a ESP application: (i) stateful operators can be bottlenecks because they are sensitive to the order of events and cannot be trivially parallelized by replication; and (ii), if failures are to be tolerated, the accumulated state of an stateful operator needs to be saved, saving this state traditionally imposes considerable performance costs. Our approach is to evaluate the use of speculation to address these two issues. For handling ordering and parallelization issues in a stateful operator, we propose a speculative approach that both reduces latency when the operator must wait for the correct ordering of the events and improves throughput when the operation in hand is parallelizable. In addition, our approach does not require that user understand concurrent programming or that he or she needs to consider out-of-order execution when writing the operations. For fault-tolerant applications, traditional approaches have imposed prohibitive performance costs due to pessimistic schemes. We extend such approaches, using speculation to mask the cost of fault tolerance.

Today, most modern online services make use of big data analytics systems to extract useful information from the raw digital data. The data normally arrives as a continuous data stream at a high speed and in huge volumes. The cost of handling this massive data can be significant. Providing interactive latency in processing the data is often impractical due to the fact that the data is growing exponentially and even faster than Moore’s law predictions. To overcome this problem, approximate computing has recently emerged as a promising solution. Approximate computing is based on the observation that many modern applications are amenable to an approximate, rather than the exact output. Unlike traditional computing, approximate computing tolerates lower accuracy to achieve lower latency by computing over a partial subset instead of the entire input data. Unfortunately, the advancements in approximate computing are primarily geared towards batch analytics and cannot provide low-latency guarantees in the context of stream processing, where new data continuously arrives as an unbounded stream. In this thesis, we design and implement approximate computing techniques for processing and interacting with high-speed and large-scale stream data to achieve low latency and efficient utilization of resources. To achieve these goals, we have designed and built the following approximate data analytics systems: • StreamApprox—a data stream analytics system for approximate computing. This system supports approximate computing for low-latency stream analytics in a transparent way and has an ability to adapt to rapid fluctuations of input data streams. In this system, we designed an online adaptive stratified reservoir sampling algorithm to produce approximate output with bounded error. • IncApprox—a data analytics system for incremental approximate computing. This system adopts approximate and incremental computing in stream processing to achieve high-throughput and low-latency with efficient resource utilization. In this system, we designed an online stratified sampling algorithm that uses self-adjusting computation to produce an incrementally updated approximate output with bounded error. • PrivApprox—a data stream analytics system for privacy-preserving and approximate computing. This system supports high utility and low-latency data analytics and preserves user’s privacy at the same time. The system is based on the combination of privacy-preserving data analytics and approximate computing. • ApproxJoin—an approximate distributed joins system. This system improves the performance of joins — critical but expensive operations in big data systems. In this system, we employed a sketching technique (Bloom filter) to avoid shuffling non-joinable data items through the network as well as proposed a novel sampling mechanism that executes during the join to obtain an unbiased representative sample of the join output. Our evaluation based on micro-benchmarks and real world case studies shows that these systems can achieve significant performance speedup compared to state-of-the-art systems by tolerating negligible accuracy loss of the analytics output. In addition, our systems allow users to systematically make a trade-off between accuracy and throughput/latency and require no/minor modifications to the existing applications.

Die Menge an verfügbaren, elektronisch auswertbaren Informationen nimmt stetig zu. Mobiltelefone mit unterschiedlichsten Sensoren, soziale Netzwerke und das Internet der Dinge sind Beispiele für Erzeuger von potentiell interessanten und verwertbaren Daten. Das Themenfeld der ereignisverarbeitenden Systeme (Event Processing – EP) bietet Technologien und Werkzeuge an, um eintreffende Daten, sog. Ereignisse, in nahezu Echtzeit zu verarbeiten. So können z.B. Muster in den Ereignissen erkannt werden. Durch die Erstellung von abgeleiteten Ereignissen können somit weitere Systemen auf diese Mustererkennung reagieren. So können u.a. zeitbasierte Funktionalitäten realisiert werden, wie z.B. das Überwachen von Aktienkursen in einem definierten Zeitraum. Im Gegensatz zu einem nachrichtenorientierten Kommunikationssystem können in EP-Anwendungen fachlich relevante Anwendungsfunktionalitäten umgesetzt werden. Die Validierung dieser Anwendungen durch Fachexperten gewinnt dadurch eine gesteigerte Bedeutung. Die abnahmetestgetriebene Entwicklung (Acceptance Test Driven Development – ATDD) ist eine Methode der agilen Softwareentwicklung und fokussiert sich auf die Integration von Fachexperten in die Erstellung und Auswertung von automatisierbaren Testfällen. Neben dem Potential der Automatisierung von manuellen Regressionstests liegt in der Methode die Möglichkeit den Wissenstransfer zwischen Entwicklern und Fachexperten zu verbessern. Die vorliegende Arbeit leistet mehrere Beiträge zur Untersuchung von ATDD im Bereich der EP-Anwendungsentwicklung. Zunächst wurden Anforderungen für eine entsprechende Werkzeugunterstützung auf Basis der Eigenschaften von EP-Anwendungen ermittelt und der Produktqualitätsklassifikationen funktionalen Eignung, Modularität und Benutzbarkeit zugeordnet. Im Rahmen einer systematischen Literaturrecherche wurden Ansätze aus der Literatur sowie die Werkzeugunterstützung der vorhandenen Produktlösungen analysiert. Dabei wurde deutlich, dass die verwandten Lösungen die identifizierten Anforderungen nicht ausreichend erfüllen. Dadurch motiviert wurde eine Testbeschreibungssprache sowie ein ausführendes, verteiltes Testsystem konzipiert und formal beschrieben. Die Testbeschreibungssprache bietet Kommandos zur produktunabhängigen Spezifikation von Testfällen an. Mit Hilfe des Testsystems ist es möglich, diese Testfälle gegen EP-Produktlösungen auszuführen. Anhand von ausgewählten Fallstudien und einer prototypischen Umsetzung des Lösungsansatzes wurde eine Validierung vorgenommen. Dabei wird ersichtlich, dass der vorgestellte Lösungsansatz den aktuellen Stand der Technik hinsichtlich funktionaler Eignung und Modularität in diesem Anwendungsbereich übersteigt. Die Benutzbarkeit wurde anhand von zwei Benutzerstudien tiefergehend untersucht. Dabei sind erste Erkenntnisse über die praktische Nutzung der Testbeschreibungssprache sowie zukünftige Fragestellungen aufgedeckt worden. In der ersten Studie wurde das Verstehen von Testfällen untersucht und dabei die automatisierbare Testbeschreibungssprache mit einer klassischen Testbeschreibungsvorlage verglichen. Hinsichtlich der Bearbeitungsdauer wurde ein signifikanter Effekt zugunsten der automatisierbaren Sprache ermittelt. Die zweite Studie betrachtet das Spezifizieren von Testfällen. Auch hier wurden Vorteile hinsichtlich der Bearbeitungsdauer aufgedeckt.

Die Konstruktion von Systemen oder Modellen, welche unter Unsicherheit und Umweltschwankungen robust arbeiten, ist eine zentrale Herausforderung sowohl im Ingenieurwesen als auch in den Naturwissenschaften. Dies ist im Design-Zentrierungsproblem formalisiert als das Finden eines Designs, welches vorgegebene Spezifikationen erfüllt und dies mit einer hohen Wahrscheinlichkeit auch noch tut, wenn die Systemparameter oder die Spezifikationen zufällig schwanken. Das Finden des Zentrums wird oft durch das Problem der Quantifizierung der Robustheit eines Systems begleitet. Hier stellen wir eine neue adaptive statistische Methode vor, um beide Probleme gleichzeitig zu lösen. Unsere Methode, Lp-Adaptation, ist durch Robustheit in biologischen Systemen und durch randomisierte Lösungen für konvexe Volumenberechnung inspiriert. Lp-Adaptation ist in der Lage, beide Probleme im allgemeinen, nicht-konvexen Fall und bei niedrigen Rechenkosten zu lösen. In dieser Arbeit beschreiben wir die Konzepte des Algorithmus und seine einzelnen Schritte. Wir testen ihn dann anhand bekannter Vergleichsfälle und zeigen seine Anwendbarkeit in elektronischen und biologischen Systemen. In allen Fällen übertrifft das vorliegende Verfahren den bisherigen Stand der Technik. Dies ermöglicht die Umformulierung von Optimierungsproblemen im Ingenieurwesen und in der Biologie als Design-Zentrierungsprobleme unter Berücksichtigung der globalen Robustheit des Systems. / The design of systems or models that work robustly under uncertainty and environmental fluctuations is a key challenge in both engineering and science. This is formalized in the design centering problem, defined as finding a design that fulfills given specifications and has a high probability of still doing so if the system parameters or the specifications randomly fluctuate. Design centering is often accompanied by the problem of quantifying the robustness of a system. Here we present a novel adaptive statistical method to simultaneously address both problems. Our method, Lp-Adaptation, is inspired by how robustness evolves in biological systems and by randomized schemes for convex volume computation. It is able to address both problems in the general, non-convex case and at low computational cost. In this thesis, we describe the concepts of the algorithm and detail its steps. We then test it on known benchmarks, and demonstrate its real-world applicability in electronic and biological systems. In all cases, the present method outperforms the previous state of the art. This enables re-formulating optimization problems in engineering and biology as design centering problems, taking global system robustness into account.

Aktuelle ökonomische Trends, wie die zunehmende Globalisierung und die wachsende Technisierung und Individualisierung vieler Konsumgüter, führen im Hinblick auf die zur Fertigung dieser Güter eingesetzte Automatisierungstechnik zu steigender Komplexität und hohen Flexibilitätsanforderungen. Ein Konzept zur Adressierung dieser Anforderungen ist die Auslegung von automatisierten Anlagen als modulares System flexibel kombinierbarer cyber-physischer Komponenten. Die namensgebende Einheit von mechatronischem Bauteil und lokaler Rechenkapazität ermöglicht Herstellern solcher Komponenten, Softwarebausteine für typische Steuer-, Bedien- oder Diagnoseaufgaben gebrauchsfertig vorzubereiten und so den (Re-)Engineeringaufwand bei der (Um-)Gestaltung des Gesamtsystems deutlich zu reduzieren. Allerdings stellt diese Vision hohe Ansprüche an die zugrundeliegende Softwarearchitektur, die von den derzeit zur Realisierung automatisierter Systeme eingesetzten Technologien nicht vollständig erfüllt werden. Das Paradigma der Agentenorientierung ist ein tragfähiger Ansatz zur Realisierung solcher lose gekoppelten verteilten Systeme und stellt durch leistungsfähige Interaktionsmechanismen sowie die enge Integration von semantischem Wissen zusätzliche Funktionalität in Aussicht: Als Agenten ausgelegte Komponenten könnten auch die logische Vernetzung untereinander während der Inbetriebnahme, nach Umrüstungen oder in Reaktion auf Betriebsstörungen teilweise selbst übernehmen. Dadurch ergeben sich Fähigkeiten wie Selbstkonfiguration und Selbstregeneration, die in der Fachliteratur unter dem Begriff Self-X zusammengefasst werden. Die fehlende Echtzeitfähigkeit, insbesondere in Bezug auf besagte Interaktionsmechanismen, hat jedoch bisher die Einsetzbarkeit von Agentensystemen in der Automatisierung limitiert und die Ausschöpfung der genannten Potentiale behindert. Deshalb wird in dieser Dissertation eine echtzeitfähige Laufzeitumgebung für Softwareagenten entworfen und anschließend die Überarbeitung bestehenden Kommunikationsmechanismen im Hinblick auf ihre Echtzeitfähigkeit vorgenommen. In diesem Kontext wird mit dem Konzept der semantischen Adressierung eine vielfältig einsetzbare Möglichkeit geschaffen, Nachrichten an ausgewählte Gruppen von Agenten mit bestimmten, semantisch beschriebenen Eigenschaften zur verschicken. Die dabei zur Wissensrepräsentation genutzten Taxonomie-Bäume bieten ein für viele Aufgabenstellungen ausreichendes Maß an Ausdrucksstärke und erlauben zudem die Verarbeitung unter harten Echtzeitbedingungen. Abschließend werden die geschaffenen Mechanismen in einem Antwortzeitmodell abgebildet, mit dem das rechtzeitige Reagieren eines Agentensystems auf lokal oder verteilt zu behandelnde Ereignisse überprüft und nachgewiesen werden kann. Damit wird ein Hauptkritikpunkt von Agentensystemen adressiert, was zu einer nachhaltigen Steigerung der Akzeptanz des Agentenparadigmas führen könnte. Während große Teile der erarbeiten Lösung als allgemeingültige Grundlagenforschung verstanden werden können, wird bei der Formulierung von Anforderungen, der Darstellung von Beispielen und der Erläuterung von Entwurfsentscheidungen immer wieder auf automatisierungstechnische Belange Bezug genommen. Außerdem wird am Ende der Arbeit eine kritische Bewertung der Ergebnisse vor dem Hintergrund eines möglichen Einsatzes in zukünftigen Automatisierungssystemen durchgeführt und damit das Gesamtbild abgerundet.

Adaptability is an intrinsic property of software systems that require adaptation to cope with dynamically changing environments. Achieving adaptability is challenging. Variability is a key solution as it enables a software system to change its behavior which corresponds to a specific need. The abstraction of variability is to manage variants, which are dynamic parts to be composed to the base system. Run-time variability realizes these variant compositions dynamically at run time to enable adaptation. Adaptation, relying on variants specified at build time, is called anticipated adaptation, which allows the system behavior to change with respect to a set of predefined execution environments. This implies the inability to solve practical problems in which the execution environment is not completely fixed and often unknown until run time. Enabling unanticipated adaptation, which allows variants to be dynamically added at run time, alleviates this inability, but it holds several implications yielding system instability such as inconsistency and run-time failures. Adaptation should be performed only when a system reaches a consistent state to avoid inconsistency. Inconsistency is an effect of adaptation happening when the system changes the state and behavior while a series of methods is still invoking. A software bug is another source of system instability. It often appears in a variant composition and is brought to the system during adaptation. The problem is even more critical for unanticipated adaptation as the system has no prior knowledge of the new variants. This dissertation aims to achieve anticipated and unanticipated adaptation. In achieving adaptation, the issues of inconsistency and software failures, which may happen as a consequence of run-time adaptation, are evidently addressed as well. Roles encapsulate dynamic behavior used to adapt players representing the base system, which is the rationale to select roles as the software system's variants. Based on the role concept, this dissertation presents three mechanisms to comprehensively address adaptation. First, a dynamic instance binding mechanism is proposed to loosely bind players and roles. Dynamic binding of roles enables anticipated and unanticipated adaptation. Second, an object-level tranquility mechanism is proposed to avoid inconsistency by allowing a player object to adapt only when its consistent state is reached. Last, a rollback recovery mechanism is proposed as a proactive mechanism to embrace and handle failures resulting from a defective composition of variants. A checkpoint of a system configuration is created before adaptation. If a specialized bug sensor detects a failure, the system rolls back to the most recent checkpoint. These mechanisms are integrated into a role-based runtime, called LyRT. LyRT was validated with three case studies to demonstrate the practical feasibility. This validation showed that LyRT is more advanced than the existing variability approaches with respect to adaptation due to its consistency control and failure handling. Besides, several benchmarks were set up to quantify the overhead of LyRT concerning the execution time of adaptation. The results revealed that the overhead introduced to achieve anticipated and unanticipated adaptation to be small enough for practical use in adaptive software systems. Thus, LyRT is suitable for adaptive software systems that frequently require the adaptation of large sets of objects.

In my dissertation, I present ATLAS — the Auto-Training Look-Ahead Scheduler. ATLAS improves service to applications with regard to two non-functional properties: timeliness and overload detection. Timeliness is an important requirement to ensure user interface responsiveness and the smoothness of multimedia operations. Overload can occur when applications ask for more computation time than the machine can offer. Interactive systems have to handle overload situations dynamically at runtime. ATLAS provides timely service to applications, accessible through an easy-to-use interface. Deadlines specify timing requirements, workload metrics describe jobs. ATLAS employs machine learning to predict job execution times. Deadline misses are detected before they occur, so applications can react early.

Event Stream Processing (ESP) is a well-established approach for low-latency data processing enabling users to quickly react to relevant situations in soft real-time. In order to cope with the sheer amount of data being generated each day and to cope with fluctuating workloads originating from data sources such as Twitter and Facebook, such systems must be highly scalable and elastic. Hence, ESP systems are typically long running applications deployed on several hundreds of nodes in either dedicated data-centers or cloud environments such as Amazon EC2. In such environments, nodes are likely to fail due to software aging, process or hardware errors whereas the unbounded stream of data asks for continuous processing. In order to cope with node failures, several fault tolerance approaches have been proposed in literature. Active replication and rollback recovery-based on checkpointing and in-memory logging (upstream backup) are two commonly used approaches in order to cope with such failures in the context of ESP systems. However, these approaches suffer either from a high resource footprint, low throughput or unresponsiveness due to long recovery times. Moreover, in order to recover applications in a precise manner using exactly once semantics, the use of deterministic execution is required which adds another layer of complexity and overhead. The goal of this thesis is to lower the overhead for fault tolerance in ESP systems. We first present StreamMine3G, our ESP system we built entirely from scratch in order to study and evaluate novel approaches for fault tolerance and elasticity. We then present an approach to reduce the overhead of deterministic execution by using a weak, epoch-based rather than strict ordering scheme for commutative and tumbling windowed operators that allows applications to recover precisely using active or passive replication. Since most applications are running in cloud environments nowadays, we furthermore propose an approach to increase the system availability by efficiently utilizing spare but paid resources for fault tolerance. Finally, in order to free users from the burden of choosing the correct fault tolerance scheme for their applications that guarantees the desired recovery time while still saving resources, we present a controller-based approach that adapts fault tolerance at runtime. We furthermore showcase the applicability of our StreamMine3G approach using real world applications and examples.