Refine
Has Fulltext
- yes (10)
Is part of the Bibliography
- yes (10)
Year of publication
- 2021 (10) (remove)
Document Type
- Doctoral Thesis (10) (remove)
Language
- English (10) (remove)
Keywords
- Algorithmische Geometrie (1)
- Algorithmus (1)
- Auto-Scaling (1)
- Außerschulische Bildung (1)
- Baseline Constrained LAMBDA (1)
- Benchmarking (1)
- Berechnungskomplexität (1)
- Beweissystem (1)
- Cloud Computing (1)
- Computer Science education (1)
Institute
- Institut für Informatik (10) (remove)
Sonstige beteiligte Institutionen
Corfu is a framework for satellite software, not only for the onboard part but also for the ground. Developing software with Corfu follows an iterative model-driven approach. The basis of the process is an engineering model. Engineers formally describe the basic structure of the onboard software in configuration files, which build the engineering model. In the first step, Corfu verifies the model at different levels. Not only syntactically and semantically but also on a higher level such as the scheduling.
Based on the model, Corfu generates a software scaffold, which follows an application-centric approach. Software images onboard consist of a list of applications connected through communication channels called topics. Corfu’s generic and generated code covers this fundamental communication, telecommand, and telemetry handling. All users have to do is inheriting from a generated class and implement the behavior in overridden methods. For each application, the generator creates an abstract class with pure virtual methods. Those methods are callback functions, e.g., for handling telecommands or executing code in threads.
However, from the model, one can not foresee the software implementation by users. Therefore, as an innovation compared to other frameworks, Corfu introduces feedback from the user code back to the model. In this way, we extend the engineering model with information about functions/methods, their invocations, their stack usage, and information about events and telemetry emission. Indeed, it would be possible to add further information extraction for additional use cases. We extract the information in two ways: assembly and source code analysis. The assembly analysis collects information about the stack usage of functions and methods.
On the one side, Corfu uses the gathered information to accomplished additional verification steps, e.g., checking if stack usages exceed stack sizes of threads. On the other side, we use the gathered information to improve the performance of onboard software. In a use case, we show how the compiled binary and bandwidth towards the ground is reducible by exploiting source code information at run-time.
The safety of future spaceflight depends on space surveillance and space traffic management, as the density of objects in Earth orbit has reached a level that requires collision avoidance maneuvers to be performed on a regular basis to avoid a mission or, in the context of human space flight, life-endangering threat. Driven by enhanced sensor systems capable of detecting centimeter-sized debris, megaconstellations and satellite miniaturization, the space debris problem has revealed many parallels to the plastic waste in our oceans, however with much less visibility to the eye. Future catalog sizes are expected to increase drastically, making it even more important to detect potentially dangerous encounters as early as possible.
Due to the limited number of monitoring sensors, continuous observation of all objects is impossible, resulting in the need to predict the orbital paths and their uncertainty via models to perform collision risk assessment and space object catalog maintenance. For many years the uncertainty models used for orbit determination neglected any uncertainty in the astrodynamic force models, thereby implicitly assuming them to be flawless descriptions of the true space environment. This assumption is known to result in overly optimistic uncertainty estimates, which in turn complicate collision risk analysis.
The keynote of this doctoral thesis is to establish uncertainty realism for low Earth orbiting satellites via a physically connected quantification of the dominant force model uncertainties, particularly multiple sources of atmospheric density uncertainty and orbital gravity uncertainty.
The resulting process noise models are subsequently integrated into classical and state of the art orbit determination algorithms. Their positive impact is demonstrated via numerical orbit determination simulations and a collision risk assessment study using all non-restricted objects in the official United States space catalogs. It is shown that the consideration of atmospheric density uncertainty and gravity uncertainty significantly improves the quality of the orbit determination and thus makes a contribution to future spaceflight safety by increasing the reliability of the uncertainty estimates used for collision risk assessment.
Educational robotics is an innovative approach to teaching and learning a variety of different concepts and skills as well as motivating students in the field of Science, Technology, Engineering, and Mathematics (STEM) education. This especially applies to educational robotics competitions such as, for example, the FIRST LEGO League, the RoboCup Junior, or the World Robot Olympiad as out-of-school and goal-oriented approach to educational robotics. These competitions have gained greatly in popularity in recent years and thousands of students participate in these competitions worldwide each year. Moreover, the corresponding technology became more accessible for teachers and students to use it in their classrooms and has arguably a high potential to impact the nature of science education at all levels. One skill, which is said to be benefitting from educational robotics, is problem solving. This thesis understands problem solving skills as engineering design skills (in contrast to scientific inquiry). Problem solving skills count as important skills as demanded by industry leaders and policy makers in the context of 21st century skills, which are relevant for students to be well-prepared for their future working life in today’s world, shaped by an ongoing process of automation, globalization, and digitalization. The overall aim of this thesis is to try to answer the question if educational robotics competitions such as the World Robot Olympiad (WRO) have a positive impact on students’ learning in terms of their problem solving skills (as part of 21st century skills). In detail, this thesis focuses on a) if students can improve their problem solving skills through participation in educational robotics competitions, b) how this skill development is accomplished, and c) the teachers’ support of their students during their learning process in the competition. The corresponding empirical studies were conducted throughout the seasons of 2018 and 2019 of the WRO in Germany. The results show overall positive effects of the participation in the WRO on students’ learning of problem solving skills. They display an increase of students’ problem solving skills, which is not moderated by other variables such as the competition’s category or age group, the students’ gender or experience, or the success of the teams at the competition. Moreover, the results indicate that students develop their problem solving skills by using a systematic engineering design process and sophisticated problem solving strategies. Lastly, the teacher’s role in the educational robotics competitions as manager and guide (in terms of the constructionist learning theory) of the students’ learning process (especially regarding the affective level) is underlined by the results of this thesis. All in all, this thesis contributes to the research gap concerning the lack of systematic evaluation of educational robotics to promote students’ learning by providing more (methodologically) sophisticated research on this topic. Thereby, this thesis follows the call for more rigorous (quantitative) research by the educational robotics community, which is necessary to validate the impact of educational robotics.
In the past two decades, there has been a trend to move from traditional television to Internet-based video services. With video streaming becoming one of the most popular applications in the Internet and the current state of the art in media consumption, quality expectations of consumers are increasing. Low quality videos are no longer considered acceptable in contrast to some years ago due to the increased sizes and resolution of devices. If the high expectations of the users are not met and a video is delivered in poor quality, they often abandon the service. Therefore, Internet Service Providers (ISPs) and video service providers are facing the challenge of providing seamless multimedia delivery in high quality. Currently, during peak hours, video streaming causes almost 58\% of the downstream traffic on the Internet. With higher mobile bandwidth, mobile video streaming has also become commonplace. According to the 2019 Cisco Visual Networking Index, in 2022 79% of mobile traffic will be video traffic and, according to Ericsson, by 2025 video is forecasted to make up 76% of total Internet traffic. Ericsson further predicts that in 2024 over 1.4 billion devices will be subscribed to 5G, which will offer a downlink data rate of 100 Mbit/s in dense urban environments.
One of the most important goals of ISPs and video service providers is for their users to have a high Quality of Experience (QoE). The QoE describes the degree of delight or annoyance a user experiences when using a service or application. In video streaming the QoE depends on how seamless a video is played and whether there are stalling events or quality degradations. These characteristics of a transmitted video are described as the application layer Quality of Service (QoS). In general, the QoS is defined as "the totality of characteristics of a telecommunications service that bear on its ability to satisfy stated and implied needs of the user of the service" by the ITU. The network layer QoS describes the performance of the network and is decisive for the application layer QoS.
In Internet video, typically a buffer is used to store downloaded video segments to compensate for network fluctuations. If the buffer runs empty, stalling occurs. If the available bandwidth decreases temporarily, the video can still be played out from the buffer without interruption. There are different policies and parameters that determine how large the buffer is, at what buffer level to start the video, and at what buffer level to resume playout after stalling. These have to be finely tuned to achieve the highest QoE for the user. If the bandwidth decreases for a longer time period, a limited buffer will deplete and stalling can not be avoided. An important research question is how to configure the buffer optimally for different users and situations. In this work, we tackle this question using analytic models and measurement studies. With HTTP Adaptive Streaming (HAS), the video players have the capability to adapt the video bit rate at the client side according to the available network capacity. This way the depletion of the video buffer and thus stalling can be avoided. In HAS, the quality in which the video is played and the number of quality switches also has an impact on the QoE. Thus, an important problem is the adaptation of video streaming so that these parameters are optimized. In a shared WiFi multiple video users share a single bottleneck link and compete for bandwidth. In such a scenario, it is important that resources are allocated to users in a way that all can have a similar QoE. In this work, we therefore investigate the possible fairness gain when moving from network fairness towards application-layer QoS fairness. In mobile scenarios, the energy and data consumption of the user device are limited resources and they must be managed besides the QoE. Therefore, it is also necessary, to investigate solutions, that conserve these resources in mobile devices. But how can resources be conserved without sacrificing application layer QoS? As an example for such a solution, this work presents a new probabilistic adaptation algorithm that uses abandonment statistics for ts decision making, aiming at minimizing the resource consumption while maintaining high QoS.
With current protocol developments such as 5G, bandwidths are increasing, latencies are decreasing and networks are becoming more stable, leading to higher QoS. This allows for new real time data intensive applications such as cloud gaming, virtual reality and augmented reality applications to become feasible on mobile devices which pose completely new research questions. The high energy consumption of such applications still remains an issue as the energy capacity of devices is currently not increasing as quickly as the available data rates. In this work we compare the optimal performance of different strategies for adaptive 360-degree video streaming.
Over the last decades, cybersecurity has become an increasingly important issue. Between 2019 and 2011 alone, the losses from cyberattacks in the United States grew by 6217%. At the same time, attacks became not only more intensive but also more and more versatile and diverse. Cybersecurity has become everyone’s concern. Today, service providers require sophisticated and extensive security infrastructures comprising many security functions dedicated to various cyberattacks. Still, attacks become more violent to a level where infrastructures can no longer keep up. Simply scaling up is no longer sufficient. To address this challenge, in a whitepaper, the Cloud Security Alliance (CSA) proposed multiple work packages for security infrastructure, leveraging the possibilities of Software-defined Networking (SDN) and Network Function Virtualization (NFV).
Security functions require a more sophisticated modeling approach than regular network functions. Notably, the property to drop packets deemed malicious has a significant impact on Security Service Function Chains (SSFCs)—service chains consisting of multiple security functions to protect against multiple at- tack vectors. Under attack, the order of these chains influences the end-to-end system performance depending on the attack type. Unfortunately, it is hard to predict the attack composition at system design time. Thus, we make a case for dynamic attack-aware SSFC reordering. Also, we tackle the issues of the lack of integration between security functions and the surrounding network infrastructure, the insufficient use of short term CPU frequency boosting, and the lack of Intrusion Detection and Prevention Systems (IDPS) against database ransomware attacks.
Current works focus on characterizing the performance of security functions and their behavior under overload without considering the surrounding infrastructure. Other works aim at replacing security functions using network infrastructure features but do not consider integrating security functions within the network. Further publications deal with using SDN for security or how to deal with new vulnerabilities introduced through SDN. However, they do not take security function performance into account. NFV is a popular field for research dealing with frameworks, benchmarking methods, the combination with SDN, and implementing security functions as Virtualized Network
Functions (VNFs). Research in this area brought forth the concept of Service Function Chains (SFCs) that chain multiple network functions after one another. Nevertheless, they still do not consider the specifics of security functions. The mentioned CSA whitepaper proposes many valuable ideas but leaves their realization open to others.
This thesis presents solutions to increase the performance of single security functions using SDN, performance modeling, a framework for attack-aware SSFC reordering, a solution to make better use of CPU frequency boosting, and an IDPS against database ransomware.
Specifically, the primary contributions of this work are:
• We present approaches to dynamically bypass Intrusion Detection Systems (IDS) in order to increase their performance without reducing the security level. To this end, we develop and implement three SDN-based approaches (two dynamic and one static).
We evaluate the proposed approaches regarding security and performance and show that they significantly increase the performance com- pared to an inline IDS without significant security deficits. We show that using software switches can further increase the performance of the dynamic approaches up to a point where they can eliminate any throughput drawbacks when using the IDS.
• We design a DDoS Protection System (DPS) against TCP SYN flood at tacks in the form of a VNF that works inside an SDN-enabled network. This solution eliminates known scalability and performance drawbacks of existing solutions for this attack type.
Then, we evaluate this solution showing that it correctly handles the connection establishment and present solutions for an observed issue. Next, we evaluate the performance showing that our solution increases performance up to three times. Parallelization and parameter tuning yields another 76% performance boost. Based on these findings, we discuss optimal deployment strategies.
• We introduce the idea of attack-aware SSFC reordering and explain its impact in a theoretical scenario. Then, we discuss the required information to perform this process.
We validate our claim of the importance of the SSFC order by analyzing the behavior of single security functions and SSFCs. Based on the results, we conclude that there is a massive impact on the performance up to three orders of magnitude, and we find contradicting optimal orders
for different workloads. Thus, we demonstrate the need for dynamic reordering.
Last, we develop a model for SSFC regarding traffic composition and resource demands. We classify the traffic into multiple classes and model the effect of single security functions on the traffic and their generated resource demands as functions of the incoming network traffic. Based on our model, we propose three approaches to determine optimal orders for reordering.
• We implement a framework for attack-aware SSFC reordering based on this knowledge. The framework places all security functions inside an SDN-enabled network and reorders them using SDN flows.
Our evaluation shows that the framework can enforce all routes as desired. It correctly adapts to all attacks and returns to the original state after the attacks cease. We find possible security issues at the moment of reordering and present solutions to eliminate them.
• Next, we design and implement an approach to load balance servers while taking into account their ability to go into a state of Central Processing Unit (CPU) frequency boost. To this end, the approach collects temperature information from available hosts and places services on the host that can attain the boosted mode the longest.
We evaluate this approach and show its effectiveness. For high load scenarios, the approach increases the overall performance and the performance per watt. Even better results show up for low load workloads, where not only all performance metrics improve but also the temperatures and total power consumption decrease.
• Last, we design an IDPS protecting against database ransomware attacks that comprise multiple queries to attain their goal. Our solution models these attacks using a Colored Petri Net (CPN).
A proof-of-concept implementation shows that our approach is capable of detecting attacks without creating false positives for benign scenarios. Furthermore, our solution creates only a small performance impact.
Our contributions can help to improve the performance of security infrastructures. We see multiple application areas from data center operators over software and hardware developers to security and performance researchers. Most of the above-listed contributions found use in several research publications.
Regarding future work, we see the need to better integrate SDN-enabled security functions and SSFC reordering in data center networks. Future SSFC should discriminate between different traffic types, and security frameworks should support automatically learning models for security functions. We see the need to consider energy efficiency when regarding SSFCs and take CPU boosting technologies into account when designing performance models as well as placement, scaling, and deployment strategies. Last, for a faster adaptation against recent ransomware attacks, we propose machine-assisted learning for database IDPS signatures.
This thesis describes the functional principle of FARN, a novel flight controller for Unmanned Aerial Vehicles (UAVs) designed for mission scenarios that require highly accurate and reliable navigation. The required precision is achieved by combining low-cost inertial sensors and Ultra-Wide Band (UWB) radio ranging with raw and carrier phase observations from the Global Navigation Satellite System (GNSS). The flight controller is developed within the scope of this work regarding the mission requirements of two research projects, and successfully applied under real conditions.
FARN includes a GNSS compass that allows a precise heading estimation even in environments where the conventional heading estimation based on a magnetic compass is not reliable. The GNSS compass combines the raw observations of two GNSS receivers with FARN’s real-time capable attitude determination. Thus, especially the deployment of UAVs in Arctic environments within the project for ROBEX is possible despite the weak horizontal component of the Earth’s magnetic field.
Additionally, FARN allows centimeter-accurate relative positioning of multiple UAVs in real-time. This enables precise flight maneuvers within a swarm, but also the execution of cooperative tasks in which several UAVs have a common goal or are physically coupled. A drone defense system based on two cooperative drones that act in a coordinated manner and carry a commonly suspended net to capture a potentially dangerous drone in mid-air was developed in conjunction with the
project MIDRAS.
Within this thesis, both theoretical and practical aspects are covered regarding UAV development with an emphasis on the fields of signal processing, guidance and control, electrical engineering, robotics, computer science, and programming of embedded systems. Furthermore, this work aims to provide a condensed reference for further research in the field of UAVs.
The work describes and models the utilized UAV platform, the propulsion system, the electronic design, and the utilized sensors. After establishing mathematical conventions for attitude representation, the actual core of the flight controller, namely the embedded ego-motion estimation and the principle control architecture are outlined. Subsequently, based on basic GNSS navigation algorithms, advanced carrier phase-based methods and their coupling to the ego-motion estimation framework are derived. Additionally, various implementation details and optimization steps of the system are described. The system is successfully deployed and tested within the two projects. After a critical examination and evaluation of the developed system, existing limitations and possible improvements are outlined.
Computer games are highly immersive, engaging, and motivating learning environments. By providing a tutorial at the start of a new game, players learn the basics of the game's underlying principles as well as practice how to successfully play the game. During the actual gameplay, players repetitively apply this knowledge, thus improving it due to repetition. Computer games also challenge players with a constant stream of new challenges which increase in difficulty over time. As a result, computer games even require players to transfer their knowledge to master these new challenges. A computer game consists of several game mechanics. Game mechanics are the rules of a computer game and encode the game's underlying principles. They create the virtual environments, generate a game's challenges and allow players to interact with the game. Game mechanics also can encode real world knowledge. This knowledge may be acquired by players via gameplay. However, the actual process of knowledge encoding and knowledge learning using game mechanics has not been thoroughly defined, yet. This thesis therefore proposes a theoretical model to define the knowledge learning using game mechanics: the Gamified Knowledge Encoding. The model is applied to design a serious game for affine transformations, i.e., GEtiT, and to predict the learning outcome of playing a computer game that encodes orbital mechanics in its game mechanics, i.e., Kerbal Space Program. To assess the effects of different visualization technologies on the overall learning outcome, GEtiT visualizes the gameplay in desktop-3D and immersive virtual reality. The model's applicability for effective game design as well as GEtiT's overall design are evaluated in a usability study. The learning outcome of playing GEtiT and Kerbal Space Program is assessed in four additional user studies. The studies' results validate the use of the Gamified Knowledge Encoding for the purpose of developing effective serious games and to predict the learning outcome of existing serious games. GEtiT and Kerbal Space Program yield a similar training effect but a higher motivation to tackle the assignments in comparison to a traditional learning method. In conclusion, this thesis expands the understanding of using game mechanics for an effective learning of knowledge. The presented results are of high importance for researches, educators, and developers as they also provide guidelines for the development of effective serious games.
This thesis is divided into two parts.
In the first part we contribute to a working program initiated by Pudlák (2017) who lists several major complexity theoretic conjectures relevant to proof complexity and asks for oracles that separate pairs of corresponding relativized conjectures. Among these conjectures are:
- \(\mathsf{CON}\) and \(\mathsf{SAT}\): coNP (resp., NP) does not contain complete sets that have P-optimal proof systems.
- \(\mathsf{CON}^{\mathsf{N}}\): coNP does not contain complete sets that have optimal proof systems.
- \(\mathsf{TFNP}\): there do not exist complete total polynomial search problems (also known as total NP search problems).
- \(\mathsf{DisjNP}\) and \(\mathsf{DisjCoNP}\): There do not exist complete disjoint NP pairs (coNP pairs).
- \(\mathsf{UP}\): UP does not contain complete problems.
- \(\mathsf{NP}\cap\mathsf{coNP}\): \(\mathrm{NP}\cap\mathrm{coNP}\) does not contain complete problems.
- \(\mathrm{P}\ne\mathrm{NP}\).
We construct several of the oracles that Pudlák asks for.
In the second part we investigate the computational complexity of balance problems for \(\{-,\cdot\}\)-circuits computing finite sets of natural numbers (note that \(-\) denotes the set difference). These problems naturally build on problems for integer expressions and integer circuits studied by Stockmeyer and Meyer (1973), McKenzie and Wagner (2007), and Glaßer et al. (2010).
Our work shows that the balance problem for \(\{-,\cdot\}\)-circuits is undecidable which is the first natural problem for integer circuits or related constraint satisfaction problems that admits only one arithmetic operation and is proven to be undecidable.
Starting from this result we precisely characterize the complexity of balance problems for proper subsets of \(\{-,\cdot\}\). These problems turn out to be complete for one of the classes L, NL, and NP.
Constraining graph layouts - that is, restricting the placement of vertices and the routing of edges to obey certain constraints - is common practice in graph drawing.
In this book, we discuss algorithmic results on two different restriction types:
placing vertices on the outer face and on the integer grid.
For the first type, we look into the outer k-planar and outer k-quasi-planar graphs, as well as giving a linear-time algorithm to recognize full and closed outer k-planar graphs Monadic Second-order Logic.
For the second type, we consider the problem of transferring a given planar drawing onto the integer grid while perserving the original drawings topology;
we also generalize a variant of Cauchy's rigidity theorem for orthogonal polyhedra of genus 0 to those of arbitrary genus.
These days, we are living in a digitalized world. Both our professional and private lives are pervaded by various IT services, which are typically operated using distributed computing systems (e.g., cloud environments). Due to the high level of digitalization, the operators of such systems are confronted with fast-paced and changing requirements. In particular, cloud environments have to cope with load fluctuations and respective rapid and unexpected changes in the computing resource demands. To face this challenge, so-called auto-scalers, such as the threshold-based mechanism in Amazon Web Services EC2, can be employed to enable elastic scaling of the computing resources. However, despite this opportunity, business-critical applications are still run with highly overprovisioned resources to guarantee a stable and reliable service operation. This strategy is pursued due to the lack of trust in auto-scalers and the concern that inaccurate or delayed adaptations may result in financial losses.
To adapt the resource capacity in time, the future resource demands must be "foreseen", as reacting to changes once they are observed introduces an inherent delay. In other words, accurate forecasting methods are required to adapt systems proactively. A powerful approach in this context is time series forecasting, which is also applied in many other domains. The core idea is to examine past values and predict how these values will evolve as time progresses. According to the "No-Free-Lunch Theorem", there is no algorithm that performs best for all scenarios. Therefore, selecting a suitable forecasting method for a given use case is a crucial task. Simply put, each method has its benefits and drawbacks, depending on the specific use case. The choice of the forecasting method is usually based on expert knowledge, which cannot be fully automated, or on trial-and-error. In both cases, this is expensive and prone to error.
Although auto-scaling and time series forecasting are established research fields, existing approaches cannot fully address the mentioned challenges: (i) In our survey on time series forecasting, we found that publications on time series forecasting typically consider only a small set of (mostly related) methods and evaluate their performance on a small number of time series with only a few error measures while providing no information on the execution time of the studied methods. Therefore, such articles cannot be used to guide the choice of an appropriate method for a particular use case; (ii) Existing open-source hybrid forecasting methods that take advantage of at least two methods to tackle the "No-Free-Lunch Theorem" are computationally intensive, poorly automated, designed for a particular data set, or they lack a predictable time-to-result. Methods exhibiting a high variance in the time-to-result cannot be applied for time-critical scenarios (e.g., auto-scaling), while methods tailored to a specific data set introduce restrictions on the possible use cases (e.g., forecasting only annual time series); (iii) Auto-scalers typically scale an application either proactively or reactively. Even though some hybrid auto-scalers exist, they lack sophisticated solutions to combine reactive and proactive scaling. For instance, resources are only released proactively while resource allocation is entirely done in a reactive manner (inherently delayed); (iv) The majority of existing mechanisms do not take the provider's pricing scheme into account while scaling an application in a public cloud environment, which often results in excessive charged costs. Even though some cost-aware auto-scalers have been proposed, they only consider the current resource demands, neglecting their development over time. For example, resources are often shut down prematurely, even though they might be required again soon.
To address the mentioned challenges and the shortcomings of existing work, this thesis presents three contributions: (i) The first contribution-a forecasting benchmark-addresses the problem of limited comparability between existing forecasting methods; (ii) The second contribution-Telescope-provides an automated hybrid time series forecasting method addressing the challenge posed by the "No-Free-Lunch Theorem"; (iii) The third contribution-Chamulteon-provides a novel hybrid auto-scaler for coordinated scaling of applications comprising multiple services, leveraging Telescope to forecast the workload intensity as a basis for proactive resource provisioning. In the following, the three contributions of the thesis are summarized:
Contribution I - Forecasting Benchmark
To establish a level playing field for evaluating the performance of forecasting methods in a broad setting, we propose a novel benchmark that automatically evaluates and ranks forecasting methods based on their performance in a diverse set of evaluation scenarios. The benchmark comprises four different use cases, each covering 100 heterogeneous time series taken from different domains. The data set was assembled from publicly available time series and was designed to exhibit much higher diversity than existing forecasting competitions. Besides proposing a new data set, we introduce two new measures that describe different aspects of a forecast. We applied the developed benchmark to evaluate Telescope.
Contribution II - Telescope
To provide a generic forecasting method, we introduce a novel machine learning-based forecasting approach that automatically retrieves relevant information from a given time series. More precisely, Telescope automatically extracts intrinsic time series features and then decomposes the time series into components, building a forecasting model for each of them. Each component is forecast by applying a different method and then the final forecast is assembled from the forecast components by employing a regression-based machine learning algorithm. In more than 1300 hours of experiments benchmarking 15 competing methods (including approaches from Uber and Facebook) on 400 time series, Telescope outperformed all methods, exhibiting the best forecast accuracy coupled with a low and reliable time-to-result. Compared to the competing methods that exhibited, on average, a forecast error (more precisely, the symmetric mean absolute forecast error) of 29%, Telescope exhibited an error of 20% while being 2556 times faster. In particular, the methods from Uber and Facebook exhibited an error of 48% and 36%, and were 7334 and 19 times slower than Telescope, respectively.
Contribution III - Chamulteon
To enable reliable auto-scaling, we present a hybrid auto-scaler that combines proactive and reactive techniques to scale distributed cloud applications comprising multiple services in a coordinated and cost-effective manner. More precisely, proactive adaptations are planned based on forecasts of Telescope, while reactive adaptations are triggered based on actual observations of the monitored load intensity. To solve occurring conflicts between reactive and proactive adaptations, a complex conflict resolution algorithm is implemented. Moreover, when deployed in public cloud environments, Chamulteon reviews adaptations with respect to the cloud provider's pricing scheme in order to minimize the charged costs. In more than 400 hours of experiments evaluating five competing auto-scaling mechanisms in scenarios covering five different workloads, four different applications, and three different cloud environments, Chamulteon exhibited the best auto-scaling performance and reliability while at the same time reducing the charged costs. The competing methods provided insufficient resources for (on average) 31% of the experimental time; in contrast, Chamulteon cut this time to 8% and the SLO (service level objective) violations from 18% to 6% while using up to 15% less resources and reducing the charged costs by up to 45%.
The contributions of this thesis can be seen as major milestones in the domain of time series forecasting and cloud resource management. (i) This thesis is the first to present a forecasting benchmark that covers a variety of different domains with a high diversity between the analyzed time series. Based on the provided data set and the automatic evaluation procedure, the proposed benchmark contributes to enhance the comparability of forecasting methods. The benchmarking results for different forecasting methods enable the selection of the most appropriate forecasting method for a given use case. (ii) Telescope provides the first generic and fully automated time series forecasting approach that delivers both accurate and reliable forecasts while making no assumptions about the analyzed time series. Hence, it eliminates the need for expensive, time-consuming, and error-prone procedures, such as trial-and-error searches or consulting an expert. This opens up new possibilities especially in time-critical scenarios, where Telescope can provide accurate forecasts with a short and reliable time-to-result.
Although Telescope was applied for this thesis in the field of cloud computing, there is absolutely no limitation regarding the applicability of Telescope in other domains, as demonstrated in the evaluation. Moreover, Telescope, which was made available on GitHub, is already used in a number of interdisciplinary data science projects, for instance, predictive maintenance in an Industry 4.0 context, heart failure prediction in medicine, or as a component of predictive models of beehive development. (iii) In the context of cloud resource management, Chamulteon is a major milestone for increasing the trust in cloud auto-scalers. The complex resolution algorithm enables reliable and accurate scaling behavior that reduces losses caused by excessive resource allocation or SLO violations. In other words, Chamulteon provides reliable online adaptations minimizing charged costs while at the same time maximizing user experience.