AMR-aware in situ indexing and scalable querying.

Query-driven analytics on scientific datasets is one of fundamental approaches for scientific discoveries. Existing studies have explored query-driven analytics on uniform resolution meshes. However, querying on adaptive mesh refinement (AMR) data has not been explored yet. As many simulations have been lately transitioning to AMR, new methods for efficient query-driven analysis on AMR data are needed. In this paper, we present the first work to support scalable AMR-aware analysis. We propose an AMR-aware hybrid index for supporting two common forms (i.e., spatial and value-based query selections) in query-driven analytics. To sustainably support future-scale analysis, we design an in situ (run-time) index building strategy with minimized performance impact to the co-located simulation. Additionally, we develop a parallel post-processing query method with an adaptive workload-balanced strategy. Our evaluation demonstrates the scalability of our in situ indexing and scalable querying methods up to 16,384 and 1,024 cores, respectively, using a Chombo-based benchmark. Compared to non-AMR-aware indexing and querying, we demonstrate up to 12.4x and 500x performance improvement, respectively.