@phdthesis{Bauer2021,
  author    = {Bauer, Andr{\´e}},
  title     = {Automated Hybrid Time Series Forecasting: Design, Benchmarking, and Use Cases},
  doi       = {10.25972/OPUS-22025},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-220255},
  school      = {Universit{\"a}t W{\"u}rzburg},
  year      = {2021},
  abstract  = {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.},
  subject      = {Zeitreihenanalyse},
  language  = {en}
}
@phdthesis{vonKistowski2019,
  author    = {von Kistowski, J{\´o}akim Gunnarsson},
  title     = {Measuring, Rating, and Predicting the Energy Efficiency of Servers},
  doi       = {10.25972/OPUS-17847},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-178478},
  school      = {Universit{\"a}t W{\"u}rzburg},
  year      = {2019},
  abstract  = {Energy efficiency of computing systems has become an increasingly important issue over the last decades. In 2015, data centers were responsible for 2\% of the world's greenhouse gas emissions, which is roughly the same as the amount produced by air travel. In addition to these environmental concerns, power consumption of servers in data centers results in significant operating costs, which increase by at least 10\% each year. To address this challenge, the U.S. EPA and other government agencies are considering the use of novel measurement methods in order to label the energy efficiency of servers. The energy efficiency and power consumption of a server is subject to a great number of factors, including, but not limited to, hardware, software stack, workload, and load level. This huge number of influencing factors makes measuring and rating of energy efficiency challenging. It also makes it difficult to find an energy-efficient server for a specific use-case. Among others, server provisioners, operators, and regulators would profit from information on the servers in question and on the factors that affect those servers' power consumption and efficiency. However, we see a lack of measurement methods and metrics for energy efficiency of the systems under consideration. Even assuming that a measurement methodology existed, making decisions based on its results would be challenging. Power prediction methods that make use of these results would aid in decision making. They would enable potential server customers to make better purchasing decisions and help operators predict the effects of potential reconfigurations. Existing energy efficiency benchmarks cannot fully address these challenges, as they only measure single applications at limited sets of load levels. In addition, existing efficiency metrics are not helpful in this context, as they are usually a variation of the simple performance per power ratio, which is only applicable to single workloads at a single load level. Existing data center efficiency metrics, on the other hand, express the efficiency of the data center space and power infrastructure, not focusing on the efficiency of the servers themselves. Power prediction methods for not-yet-available systems that could make use of the results provided by a comprehensive power rating methodology are also lacking. Existing power prediction models for hardware designers have a very fine level of granularity and detail that would not be useful for data center operators. This thesis presents a measurement and rating methodology for energy efficiency of servers and an energy efficiency metric to be applied to the results of this methodology. We also design workloads, load intensity and distribution models, and mechanisms that can be used for energy efficiency testing. Based on this, we present power prediction mechanisms and models that utilize our measurement methodology and its results for power prediction. Specifically, the six major contributions of this thesis are: We present a measurement methodology and metrics for energy efficiency rating of servers that use multiple, specifically chosen workloads at different load levels for a full system characterization. We evaluate the methodology and metric with regard to their reproducibility, fairness, and relevance. We investigate the power and performance variations of test results and show fairness of the metric through a mathematical proof and a correlation analysis on a set of 385 servers. We evaluate the metric's relevance by showing the relationships that can be established between metric results and third-party applications. We create models and extraction mechanisms for load profiles that vary over time, as well as load distribution mechanisms and policies. The models are designed to be used to define arbitrary dynamic load intensity profiles that can be leveraged for benchmarking purposes. The load distribution mechanisms place workloads on computing resources in a hierarchical manner. Our load intensity models can be extracted in less than 0.2 seconds and our resulting models feature a median modeling error of 12.7\% on average. In addition, our new load distribution strategy can save up to 10.7\% of power consumption on a single server node. We introduce an approach to create small-scale workloads that emulate the power consumption-relevant behavior of large-scale workloads by approximating their CPU performance counter profile, and we introduce TeaStore, a distributed, micro-service-based reference application. TeaStore can be used to evaluate power and performance model accuracy, elasticity of cloud auto-scalers, and the effectiveness of power saving mechanisms for distributed systems. We show that we are capable of emulating the power consumption behavior of realistic workloads with a mean deviation less than 10\% and down to 0.2 watts (1\%). We demonstrate the use of TeaStore in the context of performance model extraction and cloud auto-scaling also showing that it may generate workloads with different effects on the power consumption of the system under consideration. We present a method for automated selection of interpolation strategies for performance and power characterization. We also introduce a configuration approach for polynomial interpolation functions of varying degrees that improves prediction accuracy for system power consumption for a given system utilization. We show that, in comparison to regression, our automated interpolation method selection and configuration approach improves modeling accuracy by 43.6\% if additional reference data is available and by 31.4\% if it is not. We present an approach for explicit modeling of the impact a virtualized environment has on power consumption and a method to predict the power consumption of a software application. Both methods use results produced by our measurement methodology to predict the respective power consumption for servers that are otherwise not available to the person making the prediction. Our methods are able to predict power consumption reliably for multiple hypervisor configurations and for the target application workloads. Application workload power prediction features a mean average absolute percentage error of 9.5\%. Finally, we propose an end-to-end modeling approach for predicting the power consumption of component placements at run-time. The model can also be used to predict the power consumption at load levels that have not yet been observed on the running system. We show that we can predict the power consumption of two different distributed web applications with a mean absolute percentage error of 2.2\%. In addition, we can predict the power consumption of a system at a previously unobserved load level and component distribution with an error of 1.2\%. The contributions of this thesis already show a significant impact in science and industry. The presented efficiency rating methodology, including its metric, have been adopted by the U.S. EPA in the latest version of the ENERGY STAR Computer Server program. They are also being considered by additional regulatory agencies, including the EU Commission and the China National Institute of Standardization. In addition, the methodology's implementation and the underlying methodology itself have already found use in several research publications. Regarding future work, we see a need for new workloads targeting specialized server hardware. At the moment, we are witnessing a shift in execution hardware to specialized machine learning chips, general purpose GPU computing, FPGAs being embedded into compute servers, etc. To ensure that our measurement methodology remains relevant, workloads covering these areas are required. Similarly, power prediction models must be extended to cover these new scenarios.},
  subject      = {Benchmarking},
  language  = {en}
}
@phdthesis{Herbst2018,
  author    = {Herbst, Nikolas Roman},
  title     = {Methods and Benchmarks for Auto-Scaling Mechanisms in Elastic Cloud Environments},
  url       = {http://nbn-resolving.de/urn:nbn:de:bvb:20-opus-164314},
  school      = {Universit{\"a}t W{\"u}rzburg},
  year      = {2018},
  abstract  = {A key functionality of cloud systems are automated resource management mechanisms at the infrastructure level. As part of this, elastic scaling of allocated resources is realized by so-called auto-scalers that are supposed to match the current demand in a way that the performance remains stable while resources are efficiently used. The process of rating cloud infrastructure offerings in terms of the quality of their achieved elastic scaling remains undefined. Clear guidance for the selection and configuration of an auto-scaler for a given context is not available. Thus, existing operating solutions are optimized in a highly application specific way and usually kept undisclosed. The common state of practice is the use of simplistic threshold-based approaches. Due to their reactive nature they incur performance degradation during the minutes of provisioning delays. In the literature, a high-number of auto-scalers has been proposed trying to overcome the limitations of reactive mechanisms by employing proactive prediction methods. In this thesis, we identify potentials in automated cloud system resource management and its evaluation methodology. Specifically, we make the following contributions: We propose a descriptive load profile modeling framework together with automated model extraction from recorded traces to enable reproducible workload generation with realistic load intensity variations. The proposed Descartes Load Intensity Model (DLIM) with its Limbo framework provides key functionality to stress and benchmark resource management approaches in a representative and fair manner. We propose a set of intuitive metrics for quantifying timing, stability and accuracy aspects of elasticity. Based on these metrics, we propose a novel approach for benchmarking the elasticity of Infrastructure-as-a-Service (IaaS) cloud platforms independent of the performance exhibited by the provisioned underlying resources. We tackle the challenge of reducing the risk of relying on a single proactive auto-scaler by proposing a new self-aware auto-scaling mechanism, called Chameleon, combining multiple different proactive methods coupled with a reactive fallback mechanism. Chameleon employs on-demand, automated time series-based forecasting methods to predict the arriving load intensity in combination with run-time service demand estimation techniques to calculate the required resource consumption per work unit without the need for a detailed application instrumentation. It can also leverage application knowledge by solving product-form queueing networks used to derive optimized scaling actions. The Chameleon approach is first in resolving conflicts between reactive and proactive scaling decisions in an intelligent way. We are confident that the contributions of this thesis will have a long-term impact on the way cloud resource management approaches are assessed. While this could result in an improved quality of autonomic management algorithms, we see and discuss arising challenges for future research in cloud resource management and its assessment methods: The adoption of containerization on top of virtual machine instances introduces another level of indirection. As a result, the nesting of virtual resources increases resource fragmentation and causes unreliable provisioning delays. Furthermore, virtualized compute resources tend to become more and more inhomogeneous associated with various priorities and trade-offs. Due to DevOps practices, cloud hosted service updates are released with a higher frequency which impacts the dynamics in user behavior.},
  subject      = {Cloud Computing},
  language  = {en}
}