Heterogeneity-Aware Placement Strategies for Query Optimization

Computing hardware is changing from systems with homogeneous CPUs to systems with heterogeneous computing units like GPUs, Many Integrated Cores, or FPGAs. This trend is caused by scaling problems of homogeneous systems, where heat dissipation and energy consumption is limiting further growths in compute-performance. Heterogeneous systems provide differently optimized computing hardware, which allows different operations to be computed on the most appropriate computing unit, resulting in faster execution and less energy consumption. For database systems, this is a new opportunity to accelerate query processing, allowing faster and more interactive querying of large amounts of data. However, the current hardware trend is also a challenge as most database systems do not support heterogeneous computing resources and it is not clear how to support these systems best. In the past, mainly single operators were ported to different computing units showing great results, while missing a system wide application. To efficiently support heterogeneous systems, a systems approach for query processing and query optimization is needed. In this thesis, we tackle the optimization challenge in detail. As a starting point, we evaluate three different approaches on isolated use-cases to assess their advantages and limitations. First, we evaluate a fork-join approach of intra-operator parallelism, where the same operator is executed on multiple computing units at the same time, each execution with different data partitions. Second, we evaluate using one computing unit statically to accelerate one operator, which provides high code-optimization potential, due to this static and pre-known usage of hardware and software. Third, we evaluate dynamically placing operators onto computing units, depending on the operator, the available computing hardware, and the given data sizes. We argue that the first and second approach suffer from multiple overheads or high implementation costs. The third approach, dynamic placement, shows good performance, while being highly extensible to different computing units and different operator implementations. To automate this dynamic approach, we first propose general placement optimization for query processing. This general approach includes runtime estimation of operators on different computing units as well as two approaches for defining the actual operator placement according to the estimated runtimes. The two placement approaches are local optimization, which decides the placement locally at run-time, and global optimization, where the placement is decided at compile-time, while allowing a global view for enhanced data sharing. The main limitation of the latter is the high dependency on cardinality estimation of intermediate results, as estimation errors for the cardinalities propagate to the operator runtime estimation and placement optimization. Therefore, we propose adaptive placement optimization, allowing the placement optimization to become fully independent of cardinalities estimation, effectively eliminating the main source of inaccuracy for runtime estimation and placement optimization. Finally, we define an adaptive placement sequence, incorporating all our proposed techniques of placement optimization. We implement this sequence as a virtualization layer between the database system and the heterogeneous hardware. Our implementation approach bases on preexisting interfaces to the database system and the hardware, allowing non-intrusive integration into existing database systems. We evaluate our techniques using two different database systems and two different OLAP benchmarks, accelerating the query processing through heterogeneous execution