SIAMPerfPortDataVisAhrens2021Final.pdf [article]

James Ahrens
This talk provides an overview of performance portability approaches for the Exascale Computing Project's (ECP) visualization and data management communities. A portable approach typically provides user-oriented data structure and operator abstractions (i.e. front-end), that are translated into performant system-oriented instructions (i.e. back-end). The high-performance visualization community has essentially agreed upon the data structures and operator abstractions offered by the open-source
more » ... isualization Toolkit (VTK). VTK supports distributed-memory parallelism and is used by Ascent, an ECP in situ API, ParaView and Visit. An ECP package, VTK-m, is improving on-node parallel performance through the addition of data-parallel data structures and operators. OpenMP and Kokkos backends provide performance portability to heterogenous node architecture targets. In contrast, the high-performance data management community has a number of front-end APIs at different levels of abstractions. These include file, collective, directory, object and service-oriented approaches. Packages include MPI-IO, NetCDF, ADIOS, HDF5 and Mochi. Data management tools typically offer support for data storage and checkpoint restart. One unique aspect of these packages is the ability to use of other data management packages as backends. In addition, these packages typically have backends that directly target the unique I/O system on each high-performance supercomputer.
doi:10.6084/m9.figshare.14156576.v1 fatcat:v6htms54ung7zfiatmg54pjfce