Fletcher: A framework to integrate FPGA accelerators with Apache Arrow

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.

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):

  • Amazon EC2 F1
  • Xilinx Alveo
  • Intel OPAE
  • OpenPOWER CAPI SNAP
  • OpenPOWER OpenCAPI OC-Accel
  • Our top-level can be generated to speak AXI, so it should be easy to integrate with many existing systems. Requirements for a platform are that it provides:
    • An AXI4 (full) interface to memory holding Arrow data structures to operate on (data path), and
    • An AXI4-lite interface for MMIO to (control path) registers.

Current state

Our framework is functional, but experimental.

⚠️ This project is currently not actively worked on. ⚠️

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:

• Simple column sum example - The "Hello, World!" of Fletcher.

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:

• Host-side run-time libraries - Software libraries for run-time support on the host-machine.

Example projects

• Simple column sum example - The "Hello, World!" of Fletcher.
• String writer - Writes at over 10 GB/s to a column of strings.
• Primitive map - Maps two primitives to two primitives (simulation-only, used in testing).
• String reader - Reads some strings (simulation-only, used in testing).

External projects using Fletcher:

• Bolson - A JSON-to-Arrow IPC conversion and Pulsar publish tool.
• An FPGA-based Parquet-to-Arrow converter
• Tidre - Demo application for the Tidre regular expression accelerator
• Wiki search with Alveo - Very rapidly search a Snappy-compressed version of a Wikipedia database dump
• Regular Expression Matching
• K-Means clustering
• Posit BLAS operations
• Posit PairHMM

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