| Analytical database queries are at the core of business intelligence and decision support. To analyze the vast amounts of data available today, query execution needs to be orders of magnitude faster. Hardware advances have made a profound impact on database design and implementation. The large main memory capacity allows queries to execute exclusively in memory and shifts the bottleneck from disk access to memory bandwidth. In the new setting, to optimize query performance, databases must be aware of an unprecedented multitude of complicated hardware features. This thesis focuses on the design and implementation of highly efficient database systems by optimizing analytical query execution for all layers of modern hardware. The hardware layers include the network across multiple machines, main memory and the NUMA interconnection across multiple processors, the multiple levels of caches across multiple processor cores, and the execution pipeline within each core. For the network layer, we introduce a distributed join algorithm that minimizes the network traffic. For the memory hierarchy, we describe partitioning variants aware to the dynamics of the CPU caches and the NUMA interconnection. To improve the memory access rate of linear scans, we optimize lightweight compression variants and evaluate their trade-offs. To accelerate query execution within the core pipeline, we introduce advanced SIMD vectorization techniques generalizable across multiple operators. We evaluate our algorithms and techniques on both mainstream hardware and on many-integrated-core platforms, and combine our techniques in a new query engine design that can better utilize the features of many-core CPUs. In the era of hardware becoming increasingly parallel and datasets consistently growing in size, this thesis can serve as a compass for developing hardware-conscious databases with truly high-performance analytical query execution. |
