The scientific community is still largely wedded to a traditional HPC compute model where a task scheduler provisions bare metal nodes and the application runs directly on the node OS. Typically, the application is restricted to using the node's software libraries and compiler toolchain. In our experience this compute model causes the development environment on the HPC cluster and the user's desktop to differ, adding obstacles to software development. In addition, cluster software upgrades potentially alter the application's runtime behavior or results.
From our point of view the provisioning of virtual machines in OpenStack offers a user friendly alternative. But how does a scientific application perform on OpenStack?
Here we run a quantum chemistry application employing novel computational algorithms written in Fortran and MPI and investigate the changes necessary in the work flow and the application code and the impact of virtualization on performance and parallel scalability.
We looked at how much effort it is to run a scientific research code on an OpenStack cloud and what the performance and parallel scaling impact was.
