This talk will highlight two topics.
Topic #1: in-situ visualization of a multi-physics simulation on LLNL's Sierra Supercomputer:
Sierra is the third fastest supercomputer in the world, clocking in at 125 peta FLOPS. Adapting to the Sierra architecture poses challenges to both the simulation and visualization infrastructures. The majority of the computational power resides on the GPU accelerators and codes must effectively leverage the GPUs in order to fully utilize the machine. Additionally, computation far outpaces our ability to save data to the file system commonly used for post-hoc analysis. This limits the frequency that data can be written out to the file system In order to visualize the data with high temporal fidelity, data must be processed while in memory, requiring the simulation and visualization routines to be coupled. This coupling is known as in situ visualization
To address these challenges, modifications to both simulation and in situ visualization were needed. We recently demonstrated a successful simulation and visualization coupling using 16,384 GPUS on 4,096 nodes of Sierra. We ran Ares, a GPU-enabled multi-physics code, coupled with Ascent, an in situ visualization tool capable of utilizing the same resources as the simulation. Ares ran a 98 billion element simulation of two-fluid mixing in a spherical geometry, an important hydrodynamic problem with applications in inertial confinement fusion and other LLNL mission-critical areas. In order to leverage Sierra’s GPUs, Ares uses RAJA, a loop-based abstraction providing architecture portability
Ascent is a fly-weight in situ infrastructure being developed as part of the ALPINE project, which is part of the Exascale Computing Project (ECP). During the simulation, Ascent was used to generate images at high temporal fidelity, and selectively save out compressed portions of the data set for post-hoc analysis. VTK-m (another ECP project) provides the GPU-performant visualization and analysis algorithms for Ascent
Topic #2: adaptive in-situ:
Triggers are an important mechanism for adapting visualization, analysis, and storage actions. With this work, we describe the Ascent in situ infrastructure’s system for triggers. This system splits triggers into two components: when to per- form an action and what actions to perform. The decision for when to perform an action can be based on different types of factors, such as mesh topology, scalar fields, or performance data. The actions to perform are also varied, ranging from the traditional action of saving simulation state to disk to performing arbitrary visualizations and analyses. We also include details on the implementation and short examples demonstrating how the system can be used.
Matt Larsen is a computer scientist at Lawrence Livermore National Laboratory. He received his Ph.D. in computer science from the University of Oregon in 2016. He is the primary developer for ALPINE's ASCENT in situ library, as well as a key contributor to VTK-m and VisIt. Matt's research interests include rendering for visualization, performance modeling for visualization, and many-core architectures.