Infrastructure

MPI requires a “low-latency” infrastructure for top performance.

The figure shows an MPI application running on two different systems. One system uses a ‘fibre’ interconnect, which are fiber optics cables that run at 56GB/s. While the second system uses standard ethernet interconnect (which runs at 10GB/s). As you can see, the ethernet system starts to get hit with significant performance penalties due to the slow network.

So we would not recommend simply plugging a bunch of workstatons together in the lab using ethernet cables and trying to scale an application up. You should run these applications on high performance computing platforms. Most universities have HPC centers available to their researchers. If not, the next page will introduce you to XSEDE. Which is the organization that includes many nationally funded computing systems specifically catered to serve STEM disciplines.