Skip to content

Fletcher: A framework to integrate FPGA accelerators with Apache Arrow

License

Notifications You must be signed in to change notification settings

ahadnagy/fletcher

 
 

Repository files navigation

Fletcher logo

Fletcher: A framework to integrate FPGA accelerators with Apache Arrow

License Last Commit

Fletcher is a framework that helps to integrate FPGA accelerators with tools and frameworks that use Apache Arrow in their back-ends.

Apache Arrow specifies an in-memory format targeting large datasets and provides libraries for various languages to interface with the data in that format. Arrow prevents the need for serialization between different language run-times and provides zero-copy inter-process communication of datasets. Languages that have Arrow libraries (under development) include C, C++, Go, Java, JavaScript, Python, Ruby and Rust.

While many software projects can benefit from these advantages, hardware accelerated applications have also seen serious serialization bottlenecks. Fletcher focuses on FPGA accelerators. Through Fletcher and Arrow, interfacing efficiently between FPGA accelerator and high-level language runtimes is made available to all the supported languages.

Given a set of Arrow Schemas, Fletcher generates the following:

  • A high-performance, easy-to-use hardware interface for your accelerator kernel:
    • You provide a range indices of your Arrow RecordBatch (rather than byte address).
    • You receive (or supply) streams of the data-type specified by the schema (rather than bus words).
    • No pointer arithmetic, reordering, buffering, etc.. required - Fletcher does this for you, with high throughput (think system bandwidth).
  • A template for the accelerator kernel (to be implemented manually or using high-level synthesis)
    • You connect directly to streams of data from/to your RecordBatch rather than some memory bus interface.

Fletcher overview

Apache Arrow support

  • Fletcher currently supports reading/writing from/to multiple Arrow RecordBatches with an Arrow Schema created from any (nested) combination of:

    • Fixed-width primitives (ints, float, etc...)
    • Lists (strings, vectors, etc...)
    • Structs
    • Validity bitmaps
  • In the future we would like to support:

    • Sparse and dense unions
    • Dictionaries
    • Chunked tabular structures (Arrow::Table)

Platform support

  • Fletcher is vendor-agnostic. Our core hardware descriptions and generated code are vendor independent; we don't use any vendor IP.

  • You can simulate a Fletcher based design without a specific target platform.

  • Tested simulators include the free and open-source GHDL and the proprietary Mentor Graphics Questa/Modelsim, and Xilinx Vivado XSIM.

  • The following platforms are (partially) supported (may be work-in-progress):

Current state

Our framework is functional, but experimental and under heavy development.

Especially the development branch (which is currently our main branch) may break without notice. Some larger examples and the supported platforms are quite hard to integrate in a CI pipeline (they would take multiple days to complete and would incur significant costs for platforms such as Amazon's EC F1). For now, these larger examples and platform support resides in separate repositories (shown above) and are tested against a specific tag of this repository.

Further reading

Tutorials:

Hardware design flow:

  • Fletcher Design Generator - The design generator converts a set of Arrow Schemas to a hardware design and provides templates for your kernel.
  • Hardware library - All Fletcher core hardware components used by the design generator.

Software design flow:

Example projects

External projects using Fletcher:

Publications

If you used or studied Fletcher, please cite:

  • J. Peltenburg, J. van Straten, L. Wijtemans, L. van Leeuwen, Z. Al-Ars, and H.P. Hofstee, Fletcher: A Framework to Efficiently Integrate FPGA Accelerators with Apache Arrow*, in 29th International Conference on Field Programmable Logic and Applications (FPL) (2019) pp. 270–277.

Additional publications:

  • J. Peltenburg, J. van Straten, M. Brobbel, H.P. Hofstee, and Z. Al-Ars, Supporting Columnar In-memory Formats on FPGA: The Hardware Design of Fletcher for Apache Arrow*, in Applied Reconfigurable Computing, edited by C. Hochberger, B. Nelson, A. Koch, R. Woods, and P. Diniz (Springer International Publishing, Cham, 2019) pp. 32–47

About

Fletcher: A framework to integrate FPGA accelerators with Apache Arrow

Resources

License

Code of conduct

Stars

Watchers

Forks

Packages

No packages published

Languages

  • VHDL 63.1%
  • C++ 23.7%
  • Python 9.8%
  • Tcl 1.5%
  • C 1.2%
  • CMake 0.6%
  • Other 0.1%