Fletcher is a framework that helps to integrate FPGA accelerators with tools that use Apache Arrow in their back-ends.
Apache Arrow specifies an in-memory format for data and provides libraries to various languages to interface with the data in that format. The format prevents the need for serialization and through the libraries Arrow, zero-copy interprocess communication is possible.
While many software projects can benefit from these advantages, also hardware accelerated applications have seen serious bottlenecks, especially when working with execution frameworks that employ implicit memory mangement, because they always require some sort of serialization to take place. Therefore, if hardware accelerators such as FPGAs can make use of the Arrow format, serialization bottlenecks may be alleviated. FPGA accelerators are the main focus of this project. (If you are looking for GPGPU's, some work is being done as part of the Arrow project already, go check out the Arrow repository.)
Because the in-memory format of Arrow is standardized, it is possible to generate easy-to-use interfaces that allow hardware accelerated functions to refer to the data through row indices rather than byte addresses. In turn, the hardware accelerated function can receive one or multiple streams (depending on the type of the data) that have the actual representation desired by the designer, and will not just be some bus word with the elements of interest somewhere in between a bunch of other bytes, forcing you to think about all the alignment issues that could arise from the vast range of types that you can express through Arrow. (Not to mention to figure out all the Arrow buffers that may be involved.)
We think the two advantages of no serialization and format standardization is paramount to efficient and highly-programmable deployment of FPGA accelerators in cluster-computing frameworks and data centers. This is why we've created Fletcher; a framework to integrate FPGA accelerators with Apache Arrow.
More specifically, you provide an Arrow Schema to an interface generation framework, and you get a hardware interface that allows you to refer to data in Arrow tables as row indices, in contrast to having to deal with addresses of possibly several nested field types, all in hardware. The interface will then provide the user with ordered and aligned streams of the requested data in the type specified through the schema.
During run time, you simple pass the Arrow table to the Fletcher run-time library, and any preperation of the data will be handled by the framework. There are a few software layers that abstract platform-specific implementation details. In this way it should be able to move over both your hardware and software more easily to another platform.
Our framework is functional, but in early stages of development. It currently supports the following features:
- Any Arrow Schema created from any combination of:
- Fixed-width primitives
- Lists
- Structs
- Validity bitmaps are supported.
- Not supported (yet):
- Unions
- Dictionaries
- Chunks
- Our hardware descriptions are vendor independent; we don't use any vendor IP.
- There is an example project for the Amazon EC2 F1 and CAPI SNAP FPGA platforms.
- Our bus interface speaks AXI, so it should be easy to integrate in many existing platforms.
- Upcoming: OpenCAPI support
This work was mainly performed by people from the Computer Engineering Lab at the Delft University of Technology. Most code is written by Jeroen van Straten and Johan Peltenburg. If you would like to contribute, open issues, make pull requests, or drop us an email (j.w.peltenburg@tudelft.nl).
- How to generate a Column Reader.
- For the current state of the project, this would be your starting point in hardware design.
- How to use the host-side run-time library.
- Regular Expression matching example on Amazon EC2 F1
- Regular Expression matching example on CAPI SNAP