Register
 
Many unoptimized HPC cluster workflows result in writing large numbers of files to distributed filesystems which can create significant problems for the performance of these shared filesystems. One of the ways to alleviate this is to organize write operations inside a persistent overlay directory attached to an immutable read-only container with your scientific software. These output files will be stored separately from the base container image, and to the host filesystem an overlay appears as a single large file. I will demo running parallel OpenFOAM simulations where all output goes into overlay images, and the total number of files on the host filesystem is reduced from several million to several dozen or less.

The same approach can be used in post-processing and visualization, where you can read simulation data from multiple overlays both in serial and in parallel. In this webinar we will walk you through all stages of creating and using overlays. We assume no prior knowledge of the container technology.