Why Glasgow Raspberry Pi Cloud is different to Southampton’s Raspberry Pi Supercomputer

We have been planning our Raspberry Pi Cloud system for more than six months. However it took a long time to get funding, then acquire all the parts. At last we are good to go. However several people have told me that our project has been scooped by Southampton. I beg to differ: Here are five reasons why our Raspberry Pi Cloud is not the same as Southampton’s Pi Supercomputer.

1) Hardware

We use USB hubs to power each mini-rack. With 14 USB sockets per hub, we only need 5 hubs to power 64 Pi boards, hence only 5 mains power sockets for Pi power. Soton have 64 individual phone chargers, needing 64 mains power sockets! [our USB power solution] [embarrassing Soton photo]

2) Case design

Our Lego racks are explicitly designed to allow us to slide Pi boards into place easily. This means we can swap out dud boards and swap in new ones without having to remove any working components. Just like a real data center. Southampton’s boards seem less accessible, once in place. [our Lego rack layer] [Soton’s Lego rack layer]

3) Software stack

The Southampton system runs MPI on top of Raspian Linux. This means they can run ‘massively’ parallel computation like nbody simulations – effectively a lightweight compute cluster.  On the other hand, we run Linux containers [LXC] on top of Raspian Linux, with some IP network trickery. A ‘job’ runs in a single container instance on a single Pi node – effectively forming part of a lightweight compute cloud. Typical jobs include web servers such as lighttpd running in a container instance. Jobs can migrate between instances, either intra-Pi board or across boards.

4) Command and control infrastructure

Whereas Southampton’s MPI-based cluster seems to be autonomous, we have a pimaster server. This is able to monitor the load on all the pi boards, provide a graphical summary, and can be used to administer the system (start new jobs, migrate computation, etc). The pimaster should provide functionality similar to the ovirt tool

5) Ambition

Whereas Southampton’s MPI cluster is used to compute the value of pi (a job which can be done just as accurately and much more efficiently on a multicore x86 server), we are working on a scale model of a cloud data center. This is to give our students hands-on learning experience with data center design and management issues – making our undergraduate curriculum relevant and practical. Also this is a hot research topic. Can scale models of cloud data centers be used to do meaningful research into cloud computing?

In summary, no hard feelings: Well done to Southampton for getting the parts together faster than we did. However we feel that our project has great potential, and we are really excited about it! Please follow our blog over the next few months to see how we get on.