PowerDL – Unified In-Memory GPU Energy Profiling and Visualization Toolkit

PowerDL is a lightweight, framework-agnostic toolkit for GPU energy, power, and utilization profiling of deep learning workloads.
It supports PyTorch and TensorFlow, operates fully in-memory by default, and provides rich, publication-ready visual analytics with optional artifact export.

The toolkit is designed for:

• Interactive experimentation (no mandatory file I/O)
• Energy-aware deep learning research
• SoftwareX-style reproducible software contributions

---

Key Features

• PyTorch and TensorFlow support
• Fully in-memory profiling (no files required by default)
• Optional export for reproducibility
• High-resolution time-series sampling (NVML-based)
• Rich, selectable visual analytics
• Uniform API across frameworks
• SoftwareX-ready outputs (figures + summaries)

---

Installation

Clone the repository and install in editable mode:

git clone https://github.com/your-repo/powerdl.git
cd powerdl
pip install -e .

Requirements

• Python ≥ 3.9
• NVIDIA GPU with NVML support
• PyTorch or TensorFlow
• numpy, pandas, matplotlib

---

Quick Start

PyTorch Example

from powerdl.highlevel_torch import profile_torch

with profile_torch(out_dir=None, interval_s=0.02, verbose=1) as prof:
    prof.train(model, trainloader, optimizer, loss_fn, epochs=3)
    prof.infer_tensor(model, batch_size=256, n_samples=20000)

rep = prof.report()
rep.plot(all=True, out_dir="results/torch_figs")

TensorFlow Example

from powerdl.highlevel import profile_tf

with profile_tf(out_dir=None, device_index=0, interval_s=0.02, verbose=1) as prof:
    prof.fit(model, dataset, epochs=10)
    prof.infer_keras(model, x_inf, batch_size=512)

rep = prof.report()
rep.plot(all=True, out_dir="results/tf_figs")

---

In-Memory First Design

PowerDL always collects measurements in memory:

• GPU power (W)
• GPU utilization (%)
• Memory utilization (%)
• Time stamps
• Phase markers (training / inference / epochs)

No files are written unless explicitly requested.

Optional export:

rep.export("runs/experiment_export")

---

Report Object

After profiling, a unified Report object is returned:

rep = prof.report()

The report contains:

• samples – time-series measurements
• marks – phase and event markers
• summary – aggregated scalar metrics

---

Visualization System

PowerDL includes a registry-based visualization system.
Users can list, select, or generate all available figures.

List Available Figures

rep.list_figures()

---

Supported Figures (Complete List)

Time-Series

• power_time – GPU power vs time (phase-shaded)
• gpu_util_time – GPU utilization vs time
• mem_util_time – memory utilization vs time
• rolling_power – smoothed power trace
• rolling_util – smoothed utilization trace

Energy-Oriented

• cumulative_energy – cumulative energy over time
• energy_rate – instantaneous energy rate
• phase_energy_bar – training vs inference energy breakdown
• epoch_energy – energy per epoch (if epoch markers exist)

Distribution-Based

• power_hist – power histogram
• util_hist – utilization histogram
• power_ecdf – empirical CDF of power
• util_ecdf – empirical CDF of utilization

Phase-Aware Statistics

• power_boxplot_phase – power by phase
• util_boxplot_phase – utilization by phase
• phase_time_bar – time spent per phase

Correlation & Density

• power_util_scatter – power vs GPU utilization
• power_util_hexbin – density-based power–utilization relation

Throughput & Efficiency

• throughput_window – instantaneous throughput
• efficiency_curve – energy–performance trade-off

---

Plotting API

Plot All Figures

rep.plot(all=True, out_dir="results/figs", show=False)

Plot Selected Figures Only

rep.plot(
    figs=[
        "power_time",
        "cumulative_energy",
        "power_util_hexbin",
        "phase_energy_bar"
    ],
    out_dir="results/figs",
    show=True,
    smooth=5,
    shade_phases=True
)

Plot a Single Figure

rep.plot_one("power_ecdf", show=True)

---

Visual Quality

All figures are:

• Consistently sized
• Grid-aligned with clean axes
• Top/right spines removed
• Publication-ready
• Saved as PNG and PDF

---

Exported Artifacts (Optional)

rep.export("runs/experiment_01")

Produces:

• summary.json
• samples.csv
• marks.csv
• figures/*.png
• figures/*.pdf

This enables full reproducibility without enforcing disk I/O.

---

Design Rationale (SoftwareX)

PowerDL separates:

• Measurement
• Analysis
• Visualization

It focuses on temporal and phase-aware energy behavior, rather than only scalar energy metrics.
The toolkit supports both interactive workflows and reproducible research.

---

Typical Use Cases

• Energy-aware deep learning research
• Green AI benchmarking
• Training vs inference energy comparison
• GPU efficiency and saturation analysis
• Software reproducibility studies

---

License

MIT License

---