Query processing on OpenCL-based FPGAs : challenges and opportunities

Traditionally, FPGAs were programmed using low-level Hardware Description Languages (HDLs) like Verilog or VHDL, which made it extremely difficult to design, build and maintain systems for FPGAs. However, the recent release of OpenCL SDKs by FPGA vendors like Xilinx and Altera have significantly improved the programmability of FPGAs and have brought new research opportunities for query processing systems on FPGAs. It remains an open question whether and how we can optimize OpenCL based database engines for FPGAs. There is a gap on optimizations and tuning between OpenCL and FPGA, since OpenCL is mainly designed for parallel multi-/many-core architectures. In this paper, we attempt to answer this question under the context of pipelined query execution. For this, we first perform a detailed study of database engines on the latest generation of FPGAs. We then design an FPGA based shared pipeline query execution system (FADE) which exploits the hardware features of FPGAs and minimizes inefficiencies like the high communication reconfiguration overhead. Our experiments show that our design achieves significant performance speedup over existing approaches for pipelined query executions on FPGA. Finally, we also present the challenges and opportunities for query processing on the latest generation FPGAs.MOE (Min. of Education, S’pore)Accepted versio