Content area
Data repositories are software systems that store, retrieve, and analyze data. They are the backbone of computing infrastructure and rely on various core components, including concurrency controls and data structures. Improving their performance is essential to supporting ever-increasing computational demand. With recent trends in computer architecture, it is becoming increasingly important to consider specialized processors and how they are interconnected and can cooperate. The design of these heterogeneous systems for data repositories and their algorithmic components is the primary focus of this dissertation. More specifically, we consider utilizing central processing units (CPUs) with graphics processing units (GPUs) as co-processors and designing our systems and components with these processors in mind. Our methodology is to approach data repositories through the lens of instruction set architecture affinity (ISA affinity), or how well our algorithms and tasks map to specific processor architectures. We further consider the interconnection between processors and the additional latency and performance of moving data between co-processors. The contributions of this dissertation include the design of two systems: a cooperative CPU-GPU key-value store and a transactional system with support for heterogeneous workloads, including hybrid transactional-analytical processing through first-class support for heterogeneous architectures. We also provide an approach to semantic transactional processing through cooperative CPU-GPU processing, an architecture-agnostic framework for coalescing memory accesses in data structures for high performance, and a mapping data structure supporting linearizable point operations and range queries.
