SQL2FPGA: Automatic Acceleration of SQL Query Processing on Modern CPU-FPGA Platforms

This repository includes the code for SQL2FPGA. SQL2FPGA is a hardware-aware SQL query compilation framwork for translating and efficiently mapping SQL queries on the modern heterogeneous CPU-FPGA platforms. SQL2FPGA takes the optimized query execution plans of SQL queries from big data query processing engines (Spark SQL for now); performs hardware-aware optimizations to map query operations to FPGA accelerators (AMD/Xilinx Vitis database overlays for now); and lastly generates the deployable CPU host code and the associated FPGA accelerator configuration code.

If you find this project useful in your research, please consider citing:

@inproceedings{lu2023sql2fpga,
  title={SQL2FPGA: Automatic Acceleration of SQL Query Processing on Modern CPU-FPGA Platforms},
  author={Lu, Alec and Fang, Zhenman},
  booktitle={Proceedings of the 31st IEEE International Symposium On Field-Programmable Custom Computing Machines},
  year={2023}
}

Download SQL2FPGA

git clone https://github.com/SFU-HiAccel/SQL2FPGA.git

Environmental Setup

1. Hardware platforms (evaluated):

   • Host CPU:
     • 64-bit Ubuntu 18.04.2 LTS
   • FPGA Accelerator Design:
     • Xilinx Database Library 2020.1
   • FPGA:
     • Xilinx Alveo U280

2. Software tools (evaluated):

   • Big data query processing tool:
     • Spark 3.1.1
   • HLS tool:
     • Vitis 2020.1
     • Xilinx Runtime(XRT) 2020.1

Accelerate SQL Query Processing using SQL2FPGA

1. Import SQL2FPGA Project using IntelliJ

   • Install Scala plugin
   • Open project ...
   • Select pom.xml project file
   • Select Scala version to match 2.13
   • Build module SQL2FPGA_Top

2. Run SQL2FPGA Project on TPC-H Dataset

   • Download TPC-H Benchmark Generator

     git clone https://github.com/electrum/tpch-dbgen.git
     

   • Generate TPC-H Dataset (our evaluation covers SF1 and SF30)

     cd tpch-dbgen/
     make
     ./dbgen -s <#> 
     

   • Specifiy Query Configurations (in SQL2FPGA_Top.scala)
     • Specify Dataset File Path
       • Modify INPUT_DIR_TPCH and OUTPUT_DIR_TPCH with the generated TPC-H dataset
     • Query Specifications
       • Modify qConfig.tpch_queryNum_start and qConfig.tpch_queryNum_end to specifiy the range of queries to generate code
     • Execution Mode
       • qConfig.pure_sw_mode = 1 indicates all operators are executed on CPU and qConfig.pure_sw_mode = 0 indicates a hybrid execution mode where both CPU and FPGA accelerators are used for execution
       • qConfig.scale_factor = 1 specifies the scale factor (SF)
   • Build and Run SQL2FPGA Module
     • Build and run module SQL2FPGA_Top
   • Output:
     • CPU Host Code: test_q#.cpp
     • FPGA Configuration Code: cfgFunc_q#.hpp
     • SW Operator Function Code: q#.hpp

3. Build AMD-Xilinx's Database Accelerator Overlay Designs

   • Clone AMD-Xilinx's Vitis Libraries:

     git clone https://github.com/Xilinx/Vitis_Libraries.git
     

   • Switch to the 2020.1 branch

     cd Vitis_Libraries
     git checkout 2020.1
     

   • Build gqeJoin and gqeAggr accelerator overlay designs (this will take more than 10 hours to finish)

     cd database/L2/demos
     make -C build_join_partition/ TARGET=hw xclbin DEVICE=xilinx_u280_xdma_201920_3
     make -C build_aggr_partition/ TARGET=hw xclbin DEVICE=xilinx_u280_xdma_201920_3
     

4. Run SQL2FPGA-generated Designs on Device

   • Replace makefile at /Vitis_Libraries/database/L2/demos with the <$SQL2FPGA_HOME>/makefile
   • Move SQL2FPGA generated code to /Vitis_Libraries/database/L2/demos/host/q#/sfsql2fpga_fpga
   • Compile and execute the design at /Vitis_Libraries/database/L2/demos

     make clean
     make run TARGET=hw MODE=FPGA TB=Q# DEVICE=xilinx_u280_xdma_201920_3 TEST=SQL2FPGA
     

Now you have compeletd the entire tool flow of SQL2FPGA. Hack the code and have fun!

Known Limitations

• Whether to use hash partition logic requires manual specification in SF30.
• Device selection between CPU and FPGA overlay design is currently manually selected through empirical results.
• Output table size required manual input to achieve the best CPU-to-FPGA memory transfer time.

Authors and Contributors

SQL2FPGA is currently maintained by Alec Lu.

Besides, we thank AMD-Xilinx Vitis DB team, Prof. Jiannan Wang and Dr. Jinglin Peng from Simon Fraser University, for their insightful discussion and technical support.

Papers

More implementation details of SQL2FPGA are covered in our paper.