STEMNET Science, Technology, Engineering and Mathematics Network

In my previous post (http://networking.stemnet.org.uk/blog/punnet-raspberry-pis) I described how I'm building a cluster of Raspberry Pi computers which will help me to demonstrate how High Performance Computing (HPC) works.

Since then I have been testing the system by using "benchmark" software.  Benchmarks are computer programs that test aspects of system performance.  The particular code I've been trying is called High Performance Linpack (HPL, or just Linpack).  This is the same application that is used by HPC system adminstrators all over the world to measure the performance of their clusters, and to try to achieve a place on the Top500 list - the list of the top 500 most powerful computers in the world.

I do now have 5 fully-functioning Raspberry Pi nodes, so I was able to run Linpack tests on four of them.  By trying different combinations of supporting software, and problem size, I found the best performance I could get was 900Mflop/s - that's 900,000,000 calculations per second.

It's possible to calculate a theoretical performance peak for a computer system, although you never expect to be able to reach it in real life.  It's like having a car that can theoretically do 200mph.  Even on a racetrack you'll struggle to actually achieve 200mph - conditions would have to be absolutely perfect.  So one measure of a computer system is not just its outright performance, but its Linpack efficiency - the ratio of achieved performance to theoretical.  The best systems can deliver over 90% efficiency.

In this case, the theoretical peak is 2,800,000,000 calculations per second.  So this Raspberry Pi cluster only hits about 32% efficiency.  But any cluster computer can be affected by factors other than just its processing power.  The Raspberry Pi has a slow network connection (by HPC standards).  Also, in order to provide shared storage to all four nodes, I had to use a slow file-sharing protocol and a USB stick.  Both of these factors will have had a significant negative impact on performance.

The bottom line is that this Raspberry Pi cluster could have reached number 262 on the Top500 list - in June 1993!

Next steps: find, or create, a GRAPHICAL parallel application that will work in a classroom environment.  I need a "wow" factor!