Energy-Aware Handling of HPC Workloads: A Co-Scheduling Approach

This invited talk, delivered by a postdoctoral researcher at Helmut Schmidt University in Hamburg, presents ongoing work on energy-aware co-scheduling of HPC workloads. The approach targets the scheduler level, arguing that unlike hardware upgrades or algorithm redesign it can improve energy efficiency using existing systems. The core idea is to run a memory-bandwidth-limited application and a CPU-bound application simultaneously on the same compute node so they complement rather than compete for resources, reducing total runtime and energy. Experiments on the HSUper cluster first compare eight core-assignment scenarios using STREAM and LINPACK, distinguishing oversubscription from non-overlapping co-scheduling; the best pinning configuration achieves energy reduction with roughly 3 percent STREAM and 26 percent LINPACK performance loss. Applying this configuration to a real memory-bound computational fluid dynamics code co-scheduled with LINPACK yields 26 percent runtime savings and 20.5 percent energy reduction, but the CFD code suffers around 50 percent performance degradation, which is unacceptable for long-running simulations measured by wall-clock time per timestep. Analysis of L3 cache and memory bandwidth shows neither is saturated, yet the memory-bound application achieves lower bandwidth in simultaneous runs, indicating interference as the cause. The speaker concludes that co-scheduling can improve energy efficiency when cache and memory requirements are met, and outlines future work on deeper contention analysis, generalization across application pairs, phase-aware scheduling, and multi-node scenarios. Audience questions probe arithmetic intensity, portability to other systems, and handling applications with alternating compute- and memory-bound phases.