Integration of FPGAs in Database Management Systems: Challenges and Opportunities

摘要

In the presence of exponential growth of the data produced every day in volume, velocity, and variety, online analytical processing (OLAP) is becoming increasingly challenging. FPGAs offer hardware reconfiguration to enable query-specific pipelined and parallel data processing with the potential of maximizing throughput, speedup as well as energy and resource efficiency. However, dynamically configuring hardware accelerators to match a given OLAP query is a complex task. Furthermore, resource limitations restrict the coverage of OLAP operators. As a consequence, query optimization through partitioning the processing onto components of heterogeneous hardware/software systems seems a promising direction. While there exists work on operator placement for heterogeneous systems, it mainly targets systems combining multi-core CPUs with GPUs. However, an inclusion of FPGAs, which uniquely offer efficient and high-throughput pipelined processing at the expense of potential reconfiguration overheads, is still an open problem. We postulate that this challenge can only be met in a scalable fashion when providing a cooperative optimization between global and FPGA-specific optimizers. We demonstrate how this is addressed in two current research projects on FPGA-based query processing.