Spec Cpu2017 Speed Eval

---
name: spec-cpu2017-speed-eval
description: Evaluates CPU energy efficiency and performance under varying RAPL power caps and core counts using standard SPEC CPU 2017 benchmarks. Probes how power capping influences the trade-off between energy consumption and execution latency across memory-intensive, balanced, and compute-intensive workloads. Use when the user wants to benchmark on SPEC CPU 2017 Speed suite, or asks about evaluating this task. Reports normalized energy usage.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2506.16046
  bibtex_key: jelvani2025linuxcommand
  confidence: high
---

# spec-cpu2017-speed-eval

> How to Increase Energy Efficiency with a Single Linux Command — Jelvani et al. (2025) (arXiv:2506.16046, 2025)

## What this evaluates

Evaluates CPU energy efficiency and performance under varying RAPL power caps and core counts using standard SPEC CPU 2017 benchmarks. Probes how power capping influences the trade-off between energy consumption and execution latency across memory-intensive, balanced, and compute-intensive workloads.

## Datasets

- **SPEC CPU 2017 Speed suite** — total ?; splits: (unstated)

## Metrics

- `normalized energy usage` **(primary)** — range: ratio
  - Energy consumption divided by the baseline energy consumption (64 cores, 150W RAPL limit). Lower values indicate higher energy efficiency.
- `normalized runtime` — range: ratio
  - Execution time divided by the baseline execution time (64 cores, 150W RAPL limit). Values above 1 indicate performance loss.
- `stalled cycle ratio` — range: [0, 1]
  - Ratio of stalled CPU cycles to total CPU cycles, used to diagnose workload bottlenecks under different power configurations.

## Input / output format

**Input**: SPEC CPU 2017 Speed benchmark executables run on a server with configurable RAPL power limits and enabled core counts.

**Output**: Normalized energy usage, normalized runtime, stalled cycle ratios, and core frequency distributions for each benchmark under each configuration.

## Scoring recipe

```python
baseline_energy = run_benchmark(cores=64, rapl_w=150)
baseline_runtime = run_benchmark(cores=64, rapl_w=150)
for config in configs:
    energy = run_benchmark(config.cores, config.rapl_w)
    runtime = run_benchmark(config.cores, config.rapl_w)
    norm_energy = energy / baseline_energy
    norm_runtime = runtime / baseline_runtime
    efficiency_gain_pct = (1 - norm_energy) * 100
    perf_loss_pct = (norm_runtime - 1) * 100
```

## Common pitfalls

- Normalization baseline is strictly 64 cores and 150W RAPL limit; using a different baseline invalidates the reported percentages.
- RAPL power limits and core counts are co-varied; evaluating them independently does not match the experimental protocol.
- Stalled cycle ratio is an explanatory diagnostic metric, not the primary evaluation metric for efficiency or performance.

## Evidence (verbatim from paper)

> Each cell of the efficiency matrices contains normalized energy usage for a single run of the benchmark at a given core count and RAPL power limit while each cell of the performance matrix contains the normalized runtime. We normalize energy efficiency and run times to the default system configuration of 64 enabled cores and a RAPL power limit of 150W (marked with a blue box).

## Citation

```bibtex
@misc{jelvani2025linuxcommand,
  title={How to Increase Energy Efficiency with a Single Linux Command},
  author={Jelvani et al. (2025)},
  year={2025},
  note={arXiv:2506.16046}
}
```

- arXiv: 2506.16046