Data orientation is a common design principle in distributed data management systems. In contrast to process-oriented or transaction-oriented system designs, data-oriented architectures are based on data locality and function shipping. The tight coupling of data and processing thereon is implemented in different systems in a variety of application scenarios such as data analysis, database-as-a-service, and data management on multiprocessor systems. Data-oriented systems, i.e., systems that implement a data-oriented architecture, bundle data and operations together in tasks which are processed locally on the nodes of the distributed system. Allocation strategies, i.e., methods that decide the mapping from tasks to nodes, are core components in data-oriented systems. Good allocation strategies can lead to balanced systems while bad allocation strategies cause skew in the load and therefore suboptimal application performance and infrastructure utilization. Optimal allocation strategies are hard to find given the complexity of the systems, the complicated interactions of tasks, and the huge solution space. To ensure the scalability of data-oriented systems and to keep them manageable with hundreds of thousands of tasks, thousands of nodes, and dynamic workloads, fast and reliable allocation strategies are mandatory. In this thesis, we develop novel allocation strategies for data-oriented systems based on graph partitioning algorithms. Therefore, we show that systems from different application scenarios with different abstraction levels can be generalized to generic infrastructure and workload descriptions. We use weighted graph representations to model infrastructures with bounded and unbounded, i.e., overcommited, resources and possibly non-linear performance characteristics. Based on our generalized infrastructure and workload model, we formalize the allocation problem, which seeks valid and balanced allocations that minimize communication. Our allocation strategies partition the workload graph using solution heuristics that work with single and multiple vertex weights. Novel extensions to these solution heuristics can be used to balance penalized and secondary graph partition weights. These extensions enable the allocation strategies to handle infrastructures with non-linear performance behavior. On top of the basic algorithms, we propose methods to incorporate heterogeneous infrastructures and to react to changing workloads and infrastructures by incrementally updating the partitioning. We evaluate all components of our allocation strategy algorithms and show their applicability and scalability with synthetic workload graphs. In end-to-end--performance experiments in two actual data-oriented systems, a database-as-a-service system and a database management system for multiprocessor systems, we prove that our allocation strategies outperform alternative state-of-the-art methods.

Data orientation is a common design principle in distributed data management systems. In contrast to process-oriented or transaction-oriented system designs, data-oriented architectures are based on data locality and function shipping. The tight coupling of data and processing thereon is implemented in different systems in a variety of application scenarios such as data analysis, database-as-a-service, and data management on multiprocessor systems. Data-oriented systems, i.e., systems that implement a data-oriented architecture, bundle data and operations together in tasks which are processed locally on the nodes of the distributed system. Allocation strategies, i.e., methods that decide the mapping from tasks to nodes, are core components in data-oriented systems. Good allocation strategies can lead to balanced systems while bad allocation strategies cause skew in the load and therefore suboptimal application performance and infrastructure utilization. Optimal allocation strategies are hard to find given the complexity of the systems, the complicated interactions of tasks, and the huge solution space. To ensure the scalability of data-oriented systems and to keep them manageable with hundreds of thousands of tasks, thousands of nodes, and dynamic workloads, fast and reliable allocation strategies are mandatory. In this thesis, we develop novel allocation strategies for data-oriented systems based on graph partitioning algorithms. Therefore, we show that systems from different application scenarios with different abstraction levels can be generalized to generic infrastructure and workload descriptions. We use weighted graph representations to model infrastructures with bounded and unbounded, i.e., overcommited, resources and possibly non-linear performance characteristics. Based on our generalized infrastructure and workload model, we formalize the allocation problem, which seeks valid and balanced allocations that minimize communication. Our allocation strategies partition the workload graph using solution heuristics that work with single and multiple vertex weights. Novel extensions to these solution heuristics can be used to balance penalized and secondary graph partition weights. These extensions enable the allocation strategies to handle infrastructures with non-linear performance behavior. On top of the basic algorithms, we propose methods to incorporate heterogeneous infrastructures and to react to changing workloads and infrastructures by incrementally updating the partitioning. We evaluate all components of our allocation strategy algorithms and show their applicability and scalability with synthetic workload graphs. In end-to-end--performance experiments in two actual data-oriented systems, a database-as-a-service system and a database management system for multiprocessor systems, we prove that our allocation strategies outperform alternative state-of-the-art methods.

Pattern matching on large graphs is the foundation for a variety of application domains. The continuously increasing size of the underlying graphs requires highly parallel in-memory graph processing engines that need to consider non-uniform memory access (NUMA) and concurrency issues to scale up on modern multiprocessor systems. To tackle these aspects, a fine-grained graph partitioning becomes increasingly important. Hence, we present a classification of graph partitioning strategies and evaluate representative algorithms on medium and large-scale NUMA systems in this paper. As a scalable pattern matching processing infrastructure, we leverage a data-oriented architecture that preserves data locality and minimizes concurrency-related bottlenecks on NUMA systems. Our in-depth evaluation reveals that the optimal partitioning strategy depends on a variety of factors and consequently, we derive a set of indicators for selecting the optimal partitioning strategy suitable for a given graph and workload.