HPC systems come in all shapes and sizes; however, they tend to share a few common characteristics. They tend to be homogeneous, with a large number of identical, rack-mounted computers located in the same room and connected with very fast networking. Sometimes, especially between upgrades, an HPC cluster might be split into two processor architectures. In those cases, particular care needs to be taken in scheduling our code if those differences have important performance implications.

Most of the computers in a cluster (called compute nodes) run exactly the same operating system and the same set of software packages and are dedicated to exclusively run computations, as the name implies. Users are not typically allowed to use these machines directly.

A smaller number of nodes are special in that they are usually not as powerful as the compute nodes but do allow users to log in. They are called service nodes (or login nodes or head nodes) and are dedicated to run user...