Streaming Task Graph Scheduling for Dataflow Architectures

Explore a cutting-edge approach to scheduling task graphs on dataflow architectures in this 29-minute conference talk from #HPDC 2023. Delve into the concept of canonical task graphs and their application in streaming task graph scheduling for dataflow devices. Learn about steady-state analysis techniques and how they inform the partitioning of task graphs into temporally multiplexed components of spatially executed tasks. Discover the potential benefits of this innovative scheduling method, including increased speedup and improved device utilization compared to traditional approaches. Gain insights into the challenges and opportunities presented by dataflow accelerators in the field of high-performance computing. Follow the presentation's logical progression from introduction to conclusions, covering key topics such as spatial block partitioning, task scheduling, and real-world results on synthetic and realistic workloads.