FRμIT project meeting #1: Cambridge

The FRμIT project kick-off meeting took place on 20 June 2017 in the Computer Laboratory at the University of Cambridge. We heard about the context in which FRμIT was funded at EPSRC and the related projects, about the Mythic Beasts Raspberry Pi hosting service, initial progress from the partners, and discussed next steps for the project.

The meeting started with a talk by Gerard Parr, from the EPSRC strategic advisory board, discussing the funding context from EPSRC and how the FRμIT project fits with their strategic goals. The RIIE call, under which FRμIT was funded, grew out of the TI3 ("Towards an Intelligent Information Infrastructure") call, and was based on the observation that many countries are building national assets around compute infrastructure and other test-beds, while in the UK we have primarily local test-beds — that should be combined, to get the benefits of economies of scale and increased synergy that we see elsewhere. FRμIT combines test-beds from Glasgow, Southampton, Loughborough, and Cambridge to achieve this goal in the single board compute space. The EPSRC focus considers scale and impact, linking test-beds, security, scalability, and non-traditional environments, all of which fit FRμIT well.

The second talk of the day was by Pete Stevens from Mythic Beasts Ltd. Mythic Beasts are a Cambridge-based hosting provider, who offer a commercial Raspberry Pi-based cloud hosting service. Pete talked us through how their service works. They use Raspberry Pi 3 devices, with power-over-Ethernet HATs supplying power and connectivity via a single cable, that they host in one of their data centres. These boot from the network, using BOOTP to download a kernel and mounting the root file-system via NFS, to avoid wear on the SD card (which is the primary cause of failure otherwise), and have proven stable in production. Mythic Beasts have custom mounting for their Pis, and fit 96 Pis in 4U of rack space, with only basic forced air cooling. Their major challenge, other than cooling and physically fitting the Pis into the racks, was that the network boot infrastructure on the Pi3 requires an idle network and has precise timing requirements for the response from the BOOTP server – this led to some careful network engineering and use of VLANs to give the appearance of isolation for each Pi.

After lunch, we reviewed initial work from each of the participants. Glasgow is looking into the base testbed management infrastructure and peer-to-peer software updates, Cambridge is looking at unikernels and delay-tolerant applications, Loughborough if looking at NFV and bare metal migration, and Southampton is looking at base infrastructure and high-density cluster hardware.

We concluded by discussing the infrastructure requirements, in terms of whether nodes were part of clusters or stand-alone, the challenges of running compute nodes outside the data centre in challenged network environments (e.g., in the presence of NATs and firewalls, and in delay tolerant environments), and the applications that we might want to support on the FRμIT testbed.

Pete Stevens from Mythic Beasts has a blog post about the FRμIT project here, written based on his participation at this meeting.

Opinions expressed are my own, and do not represent those of my employers or the organisations that fund my research.