Workshop: ISAV26: In Situ AI, Analysis and Visualization
Authors: Takuma Kawamura (Japan Atomic Energy Agency) and Kazuya Shimomura and Yasuhiro Idomura (Japan Atomic Energy Agency (JAEA))
Abstract: Ensemble statistics are classically computed after every member has been written to storage. Forming them during the run instead requires gathering field data on the compute nodes, at the cost of global communication and memory copies. We present ensemble PBVR, which forms them from visualization particles instead: one set of particles circulates through the members by a ring exchange, and the mean, variance and coefficient of variation accumulate at the particle positions during batch execution. Gradients used as volume-rendering normals follow analytically from a chain-rule expansion that closes with four accumulated quantities per particle, and the target field is defined at run time by a user-supplied algebraic expression. We apply it to a 72-member CityLBM urban plume dispersion ensemble on a GPU supercomputer. Against a post hoc ParaView workflow it reaches the first image 4 to 30 times faster and transfers 60 times less data.
Back to ISAV26: In Situ AI, Analysis and Visualization Archive Listing
Back to Full Workshop Archive Listing