Workshop: 15th International Workshop on Runtime and Operating Systems for Supercomputers (ROSS 2026)
Authors: Long Cheng, Yicheng Li, Jing Chen, and Miquel Pericàs (Chalmers University of Technology and University of Gothenburg)
Abstract: In multi-GPU task-graph runtimes, task placement affects both execution time and data locality. Earliest-finish-time (EFT) placement accounts for communication required by the current task but not how placement changes data residency for later tasks, potentially increasing subsequent communication. We observe that some placement-residency conflicts expose model-level slack, allowing lower-copy alternatives without extending the modeled completion of committed work.
We propose a communication-aware task placement scheduler that evaluates lower-copy alternatives and admits them only when they preserve modeled completion relative to the EFT choice. We implement the scheduler in CUDASTF and evaluate it on five task-graph workloads on an eight-GPU NVIDIA L4 platform. On the two workloads where placement changes, the scheduler reduces modeled input-copy volume by 39-81% and improves performance by 1.18-1.31x over the EFT baseline, while retaining EFT placement when no admissible alternative exists.
Back to 15th International Workshop on Runtime and Operating Systems for Supercomputers (ROSS 2026) Archive Listing
Back to Full Workshop Archive Listing