RADDiT is a modular, real-time digital-twin platform designed to optimize datacenter energy use by linking job-level predictions, validated scheduling simulation, and grid-aware control.
RADDiT aims to evolve datacenters from passive, high-variance loads into active, intelligent participants in grid operations.
- Grid support and resilience: Aligns job scheduling and control with external energy signals (e.g., behind-the-meter PV, grid prices/demand, utility programs such as FlexConnect).
- Operational cost savings: Reduces peak power draw and demand charges.
- Sector-wide impact: Techniques are intended to scale to DOE HPC facilities, other federal labs, utilities, hyperscalers, and campus microgrids.
- Primary audiences include: DOE HPC and facility managers, grid operators and aggregators, hyperscale DC/infrastructure engineers, campus-scale energy systems and microgrid integrators.
Core innovations:
- Per-Job Power Prediction: LLM embeddings of enriched job scripts and metadata; similarity-based inference from historical job neighbors (submission-time predictions).
- Digital-Twin–Based Control: Real-time coordination of datacenter power with grid signals using validated scheduling simulations and predictive models.
- Intelligent Job Scheduling & Power Capping:
- Reprioritizes jobs in SLURM using predicted power and external energy signals (SiteFactor approach—no Slurm core changes).
- Achieved ~17 W median per-job power-prediction error (~5% relative) on CPU-exclusive jobs.
- Validated high-fidelity scheduling simulator on Kestrel job traces (alignment on throughput, wait-time distributions, and power over time).
- Demonstrated energy-aware scheduling that aligns workload with on-campus PV and a utility program (e.g., PG&E FlexConnect) via multi-objective optimization (power alignment and wait time).
This repo provides the reproducible research artifacts for the paper and project components listed above. The FastSim-based Slurm emulator used in the study is currently in NREL licensing review; corresponding inputs/outputs are provided here to ensure end-to-end reproducibility of figures and results.
| Path / file | Contents |
|---|---|
data/ |
Static datasets (job traces, simulator I/O, encrypted embeddings where required). |
scripts/ |
Executables for embedding, semantic search, prediction, and energy-aware priority. |
scripts/quickstart.py |
Quick Start #1 — sampled CPU-only pipeline on ~10k jobs. |
scripts/quickstart_embedding.py |
Quick Start #2 — end-to-end embedding → vector DB → semantic retrieval demo. |
notebooks/ |
Jupyter notebooks to regenerate paper figures. |
requirements.txt |
Python dependencies (versions pinned as needed for reproducibility). |
Note on privacy: Job scripts and metadata may contain sensitive details. This repository distributes sanitized/aggregated artifacts. All LLM embeddings were computed on-premises.
git clone https://github.com/nrel/raddit.git
cd raddit
# Create and activate a fresh Python ≥3.9 environment
python -m venv raddit_env
source raddit_env/bin/activate # Windows: raddit_env\Scripts\activate.bat
pip install --upgrade pip
pip install -r requirements.txt
Released under the BSD 3-Clause License (see LICENSE).
Covered by NREL Software Record SWR-23-34.
This work builds on prior related work available at https://github.com/NREL/eagle-jobs/, which is also covered under the same Software Record.