A Multi-Objective Scheduling Framework for Energy–Performance Optimization in Multi-GPU Task Graphs

Abstract

Multi-GPU task-graph runtimes expose scheduling information that is not available to external power controllers, including task dependencies, data residency, communication cost, device queues, and execution slack. This thesis uses that information to study energy–performance optimization inside CUDA Sequential Task Flow (CUDASTF). The proposed framework separates the problem into two runtime layers. The first layer is a HEFT-relative residual placement scheduler that keeps earliest-finish-time placement as a conservative baseline and accepts non-HEFT placements only when predicted data-movement savings justify a bounded finish-time penalty. A window-level contextual bandit adapts the aggressiveness of this residual gate. The second layer is a guarded proposal–commit DVFS controller that converts task-level frequency intents into stable per-GPU window-level frequency decisions. The evaluation uses a single-node eight-GPU NVIDIA L4 platform and five CUDASTF benchmarks. With DVFS disabled, the placement layer reduces geometric-mean latency by 9.99% relative to HEFT. Data-movement evidence shows that the strongest placement improvements coincide with substantial copy-byte reductions, while the near-neutral workload remains close to the baseline. Under fixed Bandit place ment, WinDVFS reduces geometric-mean energy by 5.32% with a 0.48% geometric-mean latency overhead and improves geometric-mean EDP by 4.87%. End to end, Bandit+WinDVFS reduces geometric-mean latency by 8.80%, energy by 11.87%, and EDP by 19.63% relative to HEFT. These results show that bounded, explainable runtime decisions can improve locality and energy efficiency without replacing HEFT with an unconstrained learned scheduler.