Bots Dvfs Eval

bots-dvfs-eval

HiDVFS: A Hierarchical Multi-Agent DVFS Scheduler for OpenMP DAG Workloads — Pivezhandi et al. (2026) (arXiv:2601.06425, 2026)

What this evaluates

Evaluates a hierarchical multi-agent reinforcement learning scheduler's ability to optimize task allocation, frequency scaling, and core selection for OpenMP DAG workloads on embedded systems. It probes the trade-off between makespan, energy consumption, and thermal constraints under real-time profiling feedback.

Datasets

• Barcelona OpenMP Tasks Suite (BOTS) — total 12; splits: test (12)

Metrics

• makespan (primary) — range: other
  • Wall-clock time to complete all tasks in the OpenMP DAG workload. Minimized as the primary objective.
• energy — range: other
  • Total power consumption integrated over the execution time of the workload. Secondary objective.

Input / output format

Input: Profiling data per execution epoch: makespan, energy, temperature, cache misses, branch misses, current core states, and task priority/affinity constraints.

Output: Scheduling decisions: core selection mask, frequency level (0–11), and task priority assignment.

Scoring recipe

def evaluate_rl_scheduler(execution_log, window=10):
    makespans = [e['makespan'] for e in execution_log]
    energies = [e['energy'] for e in execution_log]
    l10_makespan = sum(makespans[-window:]) / window
    l10_energy = sum(energies[-window:]) / window
    return {'makespan': l10_makespan, 'energy': l10_energy}

Common pitfalls

• Confusing 'sequential mode' (single application running in parallel across multiple cores) with single-threaded execution.
• Assuming energy and makespan are independent optimization targets; the paper notes that minimizing makespan indirectly reduces energy due to shorter computation times.
• Ignoring the 2 ms round-trip scheduling overhead, which is negligible compared to benchmark execution times but part of the deployment cost.

Evidence (verbatim from paper)

This subsection outlines the evaluation methodology for assessing single-agent and multi-agent Reinforcement Learning (RL) approaches, focusing on key performance metrics—makespan, energy consumption, average temperature, branch misses, and cache misses—while integrating statistical analyses to quantify the impact of critical variables: task priority, number of cores, and average frequency.

Citation

@misc{pivezhandi2026hidvfs,
  title={HiDVFS: A Hierarchical Multi-Agent DVFS Scheduler for OpenMP DAG Workloads},
  author={Pivezhandi et al. (2026)},
  year={2026},
  note={arXiv:2601.06425}
}

• arXiv: 2601.06425