Efficiency in Cluster Database Systems - Dynamic and Workload-Aware Scaling and Allocation

Database systems have been vital in all forms of data processing for a long time. In recent years, the amount of processed data has been growing dramatically, even in small projects. Nevertheless, database management systems tend to be static in terms of size and performance which makes scaling a difficult and expensive task. Because of performance and especially cost advantages more and more installed systems have a shared nothing cluster architecture. Due to the massive parallelism of the hardware programming paradigms from high performance computing are translated into data processing. Database research struggles to keep up with this trend. A key feature of traditional database systems is to provide transparent access to the stored data. This introduces data dependencies and increases system complexity and inter process communication. Therefore, many developers are exchanging this feature for a better scalability. However, explicitly managing the data distribution and data flow requires a deep understanding of the distributed system and reduces the possibilities for automatic and autonomic optimization. In this thesis we present an approach for database system scaling and allocation that features good scalability although it keeps the data distribution transparent.
The first part of this thesis analyzes the challenges and opportunities for self-scaling database management systems in cluster environments. Scalability is a major concern of Internet based applications. Access peaks that overload the application are a financial risk. Therefore, systems are usually configured to be able to process peaks at any given moment. As a result, server systems often have a very low utilization. In distributed systems the efficiency can be increased by adapting the number of nodes to the current workload. We propose a processing model and an architecture that allows efficient self-scaling of cluster database systems. In the second part we consider different allocation approaches. To increase the efficiency we present a workload-aware, query-centric model. The approach is formalized; optimal and heuristic algorithms are presented. The algorithms optimize the data distribution for local query execution and balance the workload according to the query history. We present different query classification schemes for different forms of partitioning. The approach is evaluated for OLTP and OLAP style workloads. It is shown that variants of the approach scale well for both fields of application. The third part of the thesis considers benchmarks for large, adaptive systems. First, we present a data generator for cloud-sized applications. Due to its architecture the data generator can easily be extended and configured. A key feature is the high degree of parallelism that makes linear speedup for arbitrary numbers of nodes possible. To simulate systems with user interaction, we have analyzed a productive online e-learning management system. Based on our findings, we present a model for workload generation that considers the temporal dependency of user interaction.