This year’s Department of Energy Computer Graphics Forum meeting in Pacific Grove, CA, brought together leading visualization experts from the DOE and DOD to share their experiences developing state of the art software needed to analyze results from future exascale computers. The meeting consisted mainly of invited talks spanning a broad set of topics, including: advances in display wall technology, vendor libraries that maximize performance using hardware, software updates, realistic rendering and in situ analysis.
In my talk, “Libsim Improvements to Enable Better In Situ Workflows”, I outlined the significant reductions in both data size and time spent processing the data that can be realized by extracting surfaces of interest and saving data to the XDB format. These XDB files can then be read into FieldView or XDBview.
Additional performance benefits of this workflow are gained due to the fact that subsequent post-processing does not involve reading large amounts of volume-based results. The performance benefits are magnified when the workflow is applied in situ because the data extraction can be done while data are in the solver memory as opposed to being done after writing volume data to disk. In situ workflow sidesteps the I/O bottleneck associated with writing (and later reading) large amounts of data since it restricts data to only the features of interest, which are a small subset of the original data.
My talk demonstrated Intelligent Light’s commitment to in situ and highlighted the improvements that we have made to Libsim, the VisIt in situ library, that enable it to scale to over 131K cores using the AVF-LESLIE combustion code on the Titan supercomputer at Oak Ridge National Laboratory. We have made numerous enhancements to Libsim that improve its efficiency and ability to seamlessly accept data from the host solver code. For instance, we made enhancements that permit zero-copy passing of data from the solver to Libsim when data are not organized contiguously in memory. In addition, we eliminated several bottlenecks that affected VisIt’s rendering and scaling performance on the Titan machine. We also streamlined the creation of XDB files by developing a prototype parallel XDB library based on Oak Ridge’s high performance ADIOS framework. Taken together, these improvements to Libsim and VisIt set the stage for even larger in situ runs to come and eliminate many barriers to using in situ and an extract-based workflow.