Resource Usage Benchmark Eval

resource-usage-benchmark-eval

Benchmarking Resource Usage for Efficient Distributed Deep Learning — Frey et al. (2022) (arXiv:2201.12423, 2022)

What this evaluates

This evaluation protocol measures the computational efficiency and energy consumption of distributed deep learning training runs. It probes how model architecture, dataset, and hardware constraints (GPU count, power caps, clock speeds) affect training speed and resource utilization.

Datasets

• ImageNet — total ?; splits: train (-1)
• WikiText-103 — total ?; splits: train (-1)
• QM9 — total ?; splits: train (-1)

Metrics

• training speed (primary) — range: other
  • Wall-clock time in seconds required to complete one full training epoch.
• total energy consumption — range: other
  • Total energy used during training, calculated by multiplying wall-clock time in seconds by the measured power consumption in watts.
• GPU memory utilization — range: percent
  • Percentage of GPU memory capacity used during training, monitored via nvidia-smi.
• SM utilization — range: percent
  • Percentage of streaming multiprocessor activity during training, monitored via nvidia-smi.

Input / output format

Input: Deep learning model architecture, dataset batch, and hardware configuration specifying GPU count, power cap (100/200/250W), and clock rate.

Output: Time-series logs of GPU memory and SM utilization, wall-clock time per epoch, and total energy consumption per epoch.

Scoring recipe

def calculate_metrics(wall_clock_time_s, power_w, epochs):
    training_speed = wall_clock_time_s / epochs
    energy_joules = wall_clock_time_s * power_w
    return training_speed, energy_joules

Common pitfalls

• Accuracy metrics (e.g., top-k accuracy, perplexity) are explicitly excluded from this evaluation and left for future work.
• Energy consumption is derived by multiplying wall-clock time by power draw rather than measured directly, which may mask short-term power spikes.
• Experiments vary GPU count, power caps, and clock rates simultaneously, making it difficult to isolate the impact of a single hardware constraint.

Evidence (verbatim from paper)

Throughout our experiments, for each model in Table 1, we collect time-series data on their GPU memory utilization and streaming multiprocessor (SM) utilization, training speed, and total energy consumption/expenditure throughout each of their training runs under different GPU power caps (100 W, 200 W, and 250 W), GPU clock rates (135 MHz, 735 MHz, and 1380 MHz), and number of GPUs (2, 4, 8, 16, 32, 64, etc.) aggregated on a per-epoch level.

Citation

@misc{frey2022benchmarking,
  title={Benchmarking Resource Usage for Efficient Distributed Deep Learning},
  author={Frey et al. (2022)},
  year={2022},
  note={arXiv:2201.12423}
}

• arXiv: 2201.12423