OpenPOWER Summit North America 2019

Heterogeneous computing, one of the most crucial technology to OpenPOWER systems, is extended by cache-coherent accelerator interfaces, i.e., CAPI/OpenCAPI. Database, critical workload in data-analytics, AI, etc., is key to the success of OpenPOWER ecosystem. In this proposal, a CAPI/OpenCAPI and FPGA based solution is described to accelerate database queries. In particular, an architecture with multithreading software and multi-engine FPGA logic is presented, with detailed analysis on how CAPI/OpenCAPI’s advantages of latency, bandwidth and simplified software stack can be leveraged to accelerate database. The accelerated database is based on the popular open source PostgreSQL. Query operators including regular expression matching and other general queries are implemented in FPGA logic, with 10x to 2x overall throughput improvement compared with state-of-the-art CPU version.