Preview

PDF, English Download (3MB) | Terms of use

Abstract

Database management systems are an integral part of today’s everyday life. Trends like smart applications, the internet of things, and business and social networks require applications to deal efficiently with data in various data models close to the underlying domain. Therefore, non-relational database systems provide a wide variety of database models, like graphs and documents. However, current non-relational database systems face performance challenges due to the end of Dennard scaling and therefore performance scaling of CPUs. In the meanwhile, FPGAs have gained traction as accelerators for data management. Our goal is to tackle the performance challenges of non-relational database systems with FPGA acceleration and, at the same time, address design challenges of FPGA acceleration itself. Therefore, we split this thesis up into two main lines of work: graph processing and flexible data processing. Because of the lacking benchmark practices for graph processing accelerators, we propose GraphSim. GraphSim is able to reproduce runtimes of these accelerators based on a memory access model of the approach. Through this simulation environment, we extract three performance-critical accelerator properties: asynchronous graph processing, compressed graph data structure, and multi-channel memory. Since these accelerator properties have not been combined in one system, we propose GraphScale. GraphScale is the first scalable, asynchronous graph processing accelerator working on a compressed graph and outperforms all state-of-the-art graph processing accelerators. Focusing on accelerator flexibility, we propose PipeJSON as the first FPGA-based JSON parser for arbitrary JSON documents. PipeJSON is able to achieve parsing at line-speed, outperforming the fastest, vectorized parsers for CPUs. Lastly, we propose the subgraph query processing accelerator GraphMatch which outperforms state-of-the-art CPU systems for subgraph query processing and is able to flexibly switch queries during runtime in a matter of clock cycles.