Swallow: An Energy-Transparent Many-Core Embedded Real-Time System

Swallow is a many-core embedded system based on the XMOS XS1 architecture. It is used to explore how large numbers of lightweight (500MHz x 8 threads x 64kB/core + a fast inter-core interconnect) can be effectively used. The platform is designed to allow energy monitoring of both the cores and the communication, with the aim of producing a system that can illustrate the energy use of parallel, communicating algorithms and provide intuitive relationships between software and energy consumption.

Our research objectives are:

• Investigate programming models for large-scale parallelism.
• Investigate the energy consumption in many-core systems, in particular the balance between computation and communication.
• Provide a platform for understanding the difficulties in composing many cores at the architectural level.
• Answer the question of "How much memory is needed per processing core", for a range of contemporary applications.
• Investigate what happens when a very high ratio of communication ability to storage is provided, and the impact on algorithms.
• Further development of highly parallel algorithms.

So far, three stages of work have been undertaken:

1. The creation of a 112 processor system, based on the XMOS L2 dual-core chip.
2. Addition of DRAM storage and fast, ethernet-based I/O.
3. Expansion to 480 cores.

News

1. March 2016: Hardware designs now available

   The open-source hardware designs are now available, including CAD files and PDFs.

2. November 2015: Swallow paper accepted to DATE conference

   We will be presenting an architectural overview of Swallow at DATE, in Dresden, March 2016. More details will be forthcoming.

3. Update: Swallow nOS v0.3 released on github.

4. New wiki and API documentation!

   The Swallow documentation wiki and API references are now available.

5. We have a new ethernet booting system!

   (27th March 2013): In our most recent experiment, we were able to boot 288 cores of the Swallow system via tftp in just 4.4 seconds. Previously, this would have been impossible due to a JTAG programming limit with the older tools of 127 cores and even smaller systems took minutes to program.

   You can watch a video of the blisteringly-fast load!

6. Video of 112 cores and 615 threads for only 29 Watts!

7. 16th July 2012: 16-core research board array keeps growing

8. 16th April 2012: First 16-core research boards arrive

Photos

Some photos of the Swallow boards.

Research Team

Mug shots of people who have been involved in research with Swallow.

Academic staff

• 
  Dr. Simon Hollis
• 
  Dr. Steve Kerrison
• 
  Prof. David May

Former research students

• 
  James Hanlon
• 
  James Pallister

Hardware & software sources

Each part of Swallow is available online under various open source licenses, selected as appropriate. We are collecting a list of these here for convenience.

OS

• Swallow nOS, GitHub, GPLv3.

Compilation tools

• XN file generator for arbitrary Swallow grid sizes, GitHub, XCore Open Source License.
• Swallow app example for xTIMEcomposer v13+, GitHub, XCore Open Source License.
• Swallow ethernet boot library, GitHub, XCore Open Source License.
• Swallow xlink boot library, GitHub, XCore Open Source License.
• Swallow communication library, GitHub, XCore Open Source License.
• XE Section extractor for original format binary files, GitHub, XCore Open Source License.
• Swallow manycore tools, wrapping up the existing toolchain to provide features like parallel build and TFTP loading. The swallow_v2 branch is the most recent. XCore Open Source License.

Board designs

Some expository remarks will be added in due course. In the mean time, please refer to the source, commits and branches. Hardware designs are released under the CC-BY-4.0 license, requiring attribution to the original author. The designs are collected in the Swallow Project GitHub organisation.

• Swallow slice (XMP16), the main design files. Branch V3 is the most recent designs.
• Ethernet/SDRAM board, for use in peripheral nodes, as extra storage, and for Ethernet, of course.
• Energy measurement & SDRAM, which uses the I/O of the V3 Swallow slices and can be used for both energy sampling and extra RAM.

Chip vendor tools

In addition to our own work, the chip-vendor's software development tools — xTIMEcomposer — are available under a proprietary license, and a number of additional libraries are published under open source licenses on GitHub.

Documentation

The Swallow documentation wiki and API references are now available.