Data-Aware Predictive Scheduling for Distributed-Memory Ray Tracing

Hyungman Park, Donald Fussell, Paul Navrátil

External link (DOI)
View presentation:2021-10-25T14:00:00ZGMT-0600Change your timezone on the schedule page
2021-10-25T14:00:00Z
Exemplar figure, described by caption below
Our predictive scheduling framework generalizes existing ray scheduling methods to a unified method that adapts to the characteristics of underlying data by adjusting the scheduling depth of each ray. We have path traced the distributed partitions of the Lambda2 dataset of simulated vortices using the Frontera supercomputer at the Texas Advanced Computing Center. With our predictive scheduling method, we are able to achieve a throughput of 7-33 MRays/s while sending rays across 4-128 distributed compute nodes. Lambda2 contains an aggregate of 2.3 billion unique triangles in 1024 domains extracted from a 77.3GB HDF5 file.
Abstract

None