On-demand unstructured mesh translation for reducing memory pressure during in situ analysis

When coupling two different mesh-based codes, for example with in situ analytics, the typical strategy is to explicitly copy data (deep copy) from one implementation to another, doing translation in the process. This is necessary because codes usually do not share data model interfaces or implementations. The drawback is that data duplication results in an increased memory footprint for the coupled code. An alternative strategy, which we study in this paper, is to share mesh data through on-demand, fine-grained, run-time data model translation. This saves memory, which is an increasingly scarce resource at exascale, for the increased use of in situ analysis and decreasing memory per core. We study the performance of our method compared against a deep copy with in situ analysis at scale.