Asynchronous Many Task Fast Exascale Visualizer - AMT FEVER

The simulations on which we depend must adapt to changing hardware to run at exascale. The Asynchronous Many Task (AMT) model is a viable alternative to existing techniques and has numerous advantages. In situ data visualization and analysis remains an important fixture of computing for exascale so these algorithms must be expressed in terms of AMT so they will be compatible with emerging computing runtimes and be equally able to run efficiently at exascale. JMSI Inc d.b.a. Intelligent Light proposes AMT FEVER, a reusable in situ infrastructure that can interface with AMT simulations based on Legion. Legion has demonstrated excellent scalability on National Laboratory supercomputers and is a viable programming model beyond MPI+X. AMT FEVER will explore the interface between solver and visualization where data are passed and it will contribute AMT visualization algorithms that may be composed into the host application’s task tree in a way that minimizes in situ overhead. This has the added benefit that tasks will be compatible with Legion’s application‐to‐hardware mapping capabilities, which should lend additional performance portability to the visualization algorithms. Once implemented, the AMT FEVER code will be used as a new in situ engine for VisIt for both interactive and batch (extract based) workflows. The Phase I research will address risk reduction by increasing our experience with the new data centric Legion paradigm. A running Legion environment will be established and then a mini‐app will be ported to the Legion AMT framework to develop expertise. Following that, the customary visualization pipeline will be constructed in prototype form under Legion to expose issues and establish feasibility.A solver/visualization data interface, new visualization algorithms in AMT form (including automatic runtime hardware retargeting) and collective I/O components are anticipated. Finally, benchmarking will be performed on the instrumented miniapp. AMT FEVER’s will leverage the extensive research and deep experience of the ECP software teams into visualization algorithms that are built from the ground up for the next generation exa‐scale hardware. AMT benefits industrial users of moderately sized HPC clusters that combine CPU, GPU and other hardware resources. This will accelerate the move to in situ integration within in house CFD solver codes at JMSI’s customers such as Boeing, Lockheed Martin, NASA, Cummins, Pratt & Whitney. Delivering an AMT‐based visualization toolkit in source code form will permit others to more easily build post‐processing applications for modern hardware. By integrating the AMT FEVER pipeline into VisIt, we assure that the DOE applications have access to the familiar user experience of VisIt while getting runtime performance at scale. This will also directly benefit industry participants (such as members of the ECP Industry Council) who intend to be early users of ECP hardware and software. A full one third of the council members use JMSI products.