In Situ Analysis as a Parallel I/O Problem

Abstract

This paper presents a combination of state-of-the-practice for extraction-based in situ analysis, the treatment of in situ analysis as a form of I/O, and a proposal to the community regarding the handling of in situ and I/O libraries. Extraction in situ analysis extracts subsets of data using certain techniques (e.g., isosurfacing) and saves only the extracted information. It was found that this approach, while reducing output, still poses parallel I/O issues. An initial solution was to rework ParaView Catalyst output to use ADIOS, though that only brought about minor improvements. Further aggregation, even for large-scale extraction I/O, was the primary key to solving that issue. Finally, the paper presents the position that a single library could provide I/O, in situ analysis, code coupling, and resiliency capabilities because of the overlap in steps necessary to bootstrap each.