Ecosched Hpc Scheduling Eval

ecosched-hpc-scheduling-eval

Towards Energy Efficient Co-Scheduling in HPC — Zheng et al. (2026) (arXiv:2604.17640, 2026)

What this evaluates

Evaluates an online co-scheduling framework's ability to jointly optimize GPU count selection and job packing to minimize energy consumption while maintaining performance across diverse multi-GPU HPC workloads. It probes the scheduler's capacity to handle non-linear scaling, NUMA-aware placement, and dynamic workload packing without prior knowledge of exact runtimes.

Datasets

• Multi-GPU Benchmark Suite — total 17; splits: test (17)

Metrics

• Energy Saving (primary) — range: percent
  • Percentage reduction in total energy relative to a baseline scheduler. Calculated as (E_baseline - E_method) / E_baseline.
• Makespan Improvement — range: percent
  • Percentage reduction in total completion time relative to a baseline. Calculated as (T_baseline - T_method) / T_baseline.
• EDP Saving — range: percent
  • Percentage reduction in end-to-end Energy-Delay Product (EDP = Energy × Makespan) relative to a baseline. Calculated as (EDP_baseline - EDP_method) / EDP_baseline.
• Performance Loss — range: percent
  • Percentage increase in runtime relative to solo execution with the performance-optimal GPU count. Calculated as (T_method - T_optimal_solo) / T_optimal_solo.

Input / output format

Input: A finite window of multi-GPU jobs specifying application type, required GPU count, and NUMA constraints, along with system topology (CPU/GPU count, NUMA nodes).

Output: A schedule assigning each job a GPU count, NUMA node placement, and execution start time.

Scoring recipe

def score(gold, predictions):
    energy_saving = (gold['baseline_energy'] - predictions['method_energy']) / gold['baseline_energy']
    makespan_improvement = (gold['baseline_makespan'] - predictions['method_makespan']) / gold['baseline_makespan']
    edp_saving = (gold['baseline_edp'] - predictions['method_edp']) / gold['baseline_edp']
    performance_loss = (predictions['method_runtime'] - gold['baseline_runtime']) / gold['baseline_runtime']
    return energy_saving, makespan_improvement, edp_saving, performance_loss

Common pitfalls

• Confusing the offline Oracle baseline (which assumes perfect knowledge of runtime and power profiles) with a realistic online scheduler, leading to overly optimistic performance gap expectations.
• Evaluating only on a single workload type instead of the full 17-workload suite, which masks cross-job interference and non-linear scaling effects.
• Ignoring NUMA-aware placement constraints when measuring energy or performance, which inflates baseline metrics and invalidates the scheduling comparison.

Evidence (verbatim from paper)

EcoSched’s objective is to minimize total energy to completion of all workloads while meeting the performance loss bound, hence we use the following metrics: Energy Saving: percentage reduction in total energy relative to a baseline. Makespan Improvement: the percentage reduction in total completion time. EDP Saving: the percentage reduction in end-to-end EDP. Performance Loss: percentage increase in runtime relative to solo execution with the performance-optimal GPU count.

Citation

@misc{zheng2026ecosched,
  title={Towards Energy Efficient Co-Scheduling in HPC},
  author={Zheng et al. (2026)},
  year={2026},
  note={arXiv:2604.17640}
}

• arXiv: 2604.17640