Porting ParaFEM to the Raspberry Pi

We have a guest article from Shane Balani at the University of Bristol. Shane has been working with Anton Schterenlikht, evaluating ParaFEM on a DIY Raspberry Pi cluster. Shane's article follows below.

The benefits of a supercomputer are plain to see when running complex FEA jobs, cutting computational time down from weeks and months to a matter of hours. A supercomputer, however, is not the most common of household objects, but with the introduction of versatile and cheap computers, such as the Raspberry Pi, a cluster can be built for under £150 that can utilise the ParaFEM library and run intensive finite element analysis jobs in parallel.

The cluster consists of 4 Raspberry Pis, each working at a modest 1ghz, with a network hub to allow communication between nodes. While 4 was chosen to prove a concept, this can be scaled up much more, adding further computing power.

So what about performance? A cluster that can be built for a fraction of the millions it takes to construct a modern supercomputer, can never be expected to offer the same speed, however it's remarkable efficiency offers a big insight into the potential that exists. Running the driver program P122 on ARCHER (the UK's national supercomputer) using 4 cores, the analysis completes in 20 seconds. This can be reduced to 6 seconds by adding more cores. The 4 Pis can do the same job in 2600 seconds. A much longer time, but as shown in our tests, the speed up offered by adding more processors shows significant decreases in time taken. 1 Pi requires around 8400 seconds to run the same job.

The question of whether it’s worth adding more processors exists however, and considering the energy efficiency of the Pis, shows the huge potential that supercomputing with ARM architecture offers. A measure of performance per watt of power used reveals that the Pis offer an impressive 141 MFLOPS/watt. The Green 500 list , which ranks the top 500 most energy efficient supercomputers in the world ranges from the number 1 system with 5200MFLOPS/watt to 22MFLOPS/watt at number 500.

Our £150 Raspberry Pi cluster would rank at 440th!