martinkiefer / scotch

Scotch: Generating FPGA-Accelerators for Sketching Algorithms

Scotch

Scotch is a framework for generating optimized FPGA-accelerators for sketching algorithms.

It provides three core features:

• ScotchDSL: A domain specific language + programming models to describe a variety of sketching algorithms
• Code Generators: ScotchDSL specifications are automatically translated into a VHDL architecture containing all necessary components to perform sketching.
• Auto-Tune: An automated tuning algorithms optimizes the size of the sketch summary with respect to provided constraints and resources on the FPGA

Scotch has a corresponding publication in PVLDB:

Martin Kiefer, Ilias Poulakis, Sebastian Breß, and Volker Markl. Scotch: Generating FPGA-Accelerators for Sketching at Line Rate. PVLDB, 14(3), 2021.

Requirements

• A recent Linux operating system. For the code generator in isolation, OSX will do either.
• Python 3.6+ is required. We used Python 3.6.
• The following Python modules are required (used version in brackets): antlr4-python3-runtime (4.7.2), numpy (1.170), pandas (0.25.0).
• The ANTLR4 parser generator is required to generate the ScotchDSL parser (4.7.2). A setup script is provided for your convenience that downloads the appropriate jar and generates the parser. See README in the ScotchDSL folder.
• An FPGA toolchain is required. Intel FPGAs with Quartus Prime are supported. We tested with Quartus Prime Pro 19.3 and Quartus Prime 19.1. Thus all *10 and *V product lines should be supported. Furthermore, we support support Xilinx /Vivado. We tested with Version 2020.1.
• CPU baselines require GCC and Boost. We used GCC 7 and Boost 1.53.
• GPU baselines require CUDA and Boost. We used CUDA 10.2 with GCC 6 and Boost 1.53.

Project Structure

ScotchDSL: Contains all code generator files and implementations of various algorithms in ScotchDSL.

Autotune: Contains all files regarding automated tuning.

Baselines: CPU and GPU baseline implementations.

Sketches: ScotchDSL implementations of sketching algorithms.

IO-Controllers: I/O controller examples.

Util: Miscallaneous little helpers. Contains scripts for scaling experiments.

Our reproducibility/availability wiki page can help to get you started.